See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/395796123

A Pathway Study for Future Humanoid Standards

Technical Report · September 2025

DOI: 10.13140/RG.2.2.27892.21122

CITATIONS READS

0 1,099

16 authors, including:

Aaron Prather Ingo Keller

ASTM International Heriot-Watt University
2 PUBLICATIONS 0 CITATIONS 26 PUBLICATIONS 1,052 CITATIONS

SEE PROFILE SEE PROFILE

Brandon James DeHart

University of Waterloo
18 PUBLICATIONS 41 CITATIONS

SEE PROFILE

All content following this page was uploaded by Aaron Prather on 24 September 2025.

The user has requested enhancement of the downloaded file.

A Pathway Study For
Future Humanoid Standards
IEEE Humanoid Study Group
TABLE OF CONTENTS
1. Building the Framework for Humanoid Robotics Standards – p. 3
2. The Humanoid Robot Market – p. 6
3. Unique Risks Humanoids Bring & Why Standards Must Evolve – p. 11
4. The Classification of Humanoid Robots – p. 18
5. The Role of Stability – p. 33
6. Human–Robot Interaction (HRI) – p. 59
7. Report Summary – p. 88
8. List of Contributors – p. 92
9. Bibliography - p. 93
 Introduction: Building the Framework for Humanoid
Robotics Standards

Author: Aaron Prather, IEEE Humanoid Study Group Lead

Humanoid robots occupy a unique and highly anticipated space in the robotics
landscape. Unlike other automation systems, they promise to operate in
environments designed for humans, performing tasks as diverse as industrial
assembly, healthcare support, and public-facing services. Their appeal is
obvious: a single robotic form that can, in theory, adapt to almost any setting.

Yet, that promise largely remains unfulfilled. The reality is that humanoids face
much greater challenges than most robotic systems, not only technically but
also in how they are evaluated, certified, and trusted. The current standards
framework is not designed for them. Most existing robot standards assume
fixed or statically stable systems and do not consider the dynamic, inherently
unstable nature of a humanoid’s locomotion. Nor do they fully address the
complex ways these machines interact with people, not just physically, but
socially and psychologically.

This report seeks to bridge that gap. It is not a set of final answers, but rather a
framework of findings and recommendations that can guide the next phase of
development for humanoid standards. The analysis draws on three critical
themes—classification, stability, and human-robot interaction—each of which is
deeply interconnected and essential to moving beyond pilot programs toward
widespread deployment.

Why This Matters Now

For those developing humanoids, these findings are a roadmap to broader

acceptance. Manufacturers need clear criteria to prove that their robots are safe
and effective. Customers need assurance that the robots they adopt will
function reliably in their intended environments. Regulators require standards
that are precise enough to be applied consistently, yet flexible enough to
support innovation.

3 of 95
For Standards Development Organizations (SDOs), the need is even more
pressing. Humanoids are not just another class of mobile robots; they combine
characteristics from nearly every existing category. Without a unified approach
to classification and risk assessment, different SDOs risk producing fragmented
or conflicting requirements that slow adoption and erode public trust.

The timing is critical. Investment in humanoids is accelerating globally, with

manufacturers already piloting systems in factories, logistics centers, and
healthcare facilities. Governments are beginning to explore policies for robots
operating in public spaces. Without dedicated standards, the market will either
move forward without adequate safety assurances or remain stagnant due to
uncertainty.

A Coordinated Path Forward

This report is written not only for engineers and researchers, but also for SDO
members who will turn these ideas into actionable standards. It is designed to
show where existing standards can be extended, where entirely new ones are
required, and how organizations can collaborate.

• Classification is presented as the foundational shared taxonomy to identify

how humanoids relate to other robotic systems, what capabilities it
should be expected to have, and which standards apply.
• Stability is identified as the critical bottleneck—the area where
performance metrics and safety validation are most urgently needed.
• Human-Robot Interaction is explored as both a safety and societal issue,
requiring new guidance for physical interaction, workflow impacts, and
even perceived safety.

Taken together, these elements provide a structured approach to answering the

questions regulators, customers, and the public are already asking: Which
humanoids can we trust, and in which environments?

4 of 95
The Value of This Report

The findings outlined here are valuable to two groups in particular:

• For innovators and manufacturers, they provide insight into how to design
humanoids that can be certified and deployed in diverse human
environments. The recommendations on classification, stability testing,
and interaction guidelines will help align engineering priorities with future
regulatory expectations.
• For Standards Development Organizations, this report offers a starting
point for coordination. It highlights where ASTM, IEEE, and ISO efforts can
intersect—ASTM leading on test methods, IEEE on performance metrics,
ISO on safety thresholds—and why these must evolve together rather
than in isolation.

If humanoids are to progress from prototypes to mainstream tools, their success

will depend as much on shared standards as on technical breakthroughs.
Without clear, harmonized guidelines, deployment will remain slow,
inconsistent, and limited to tightly controlled environments. With them,
manufacturers and regulators can move forward with confidence, knowing that
safety, performance, and trust have been addressed in a systematic, evidence-
based way.

Looking Ahead

The chapters that follow dive deeper into each theme, presenting the details
behind these findings and offering concrete recommendations for moving
forward. While the path to full standardization will take time and require close
collaboration among multiple SDOs, the framework is ready to be built. For
those shaping the future of robotics, including engineers, researchers, and
standards professionals, this report serves as both a guide and a call to action.
The decisions made today about classification, stability, and human-robot
interaction will determine whether humanoids remain a niche technology or
become trusted, integrated tools in the spaces where we live and work.

5 of 95
 The Humanoid Robot Market

Authors: Mike Oitzman, Aaron Prather, and Robert Little

The humanoid robot market presents a fascinating enigma. Despite the

optimistic projections of venture capitalists and the fervent visions of founders,
it's a space that is simultaneously immense, stubbornly difficult to categorize,
and always seemingly five years in the future. The automation market is also
segmented into three distinct spaces: industrial, public/commercial, and home.
Each of these spaces has unique safety and feature requirements. The bottom
line is that a robot should never harm a human.

Science fiction has long been captivated by humanoid robots – machines

designed and built to emulate human appearance, setting them apart from
robots with specific industrial functions. Humanoid robots, by their nature, are
general-purpose machines, distinct from single-purpose robots deployed into
commercial and industrial settings. Industrial robots are often optimized for
speed and/or accuracy, able to perform tasks that humans can not.

It could be said that humanoids are designed to “displace” rather than “replace”
humans. People can deploy general-purpose humanoid robots to perform
menial tasks while we attend to other, “higher-value” tasks. By contrast,
industrial robots are designed to move faster, move more precisely, and lift
heavier payloads than humans can. Industrial automation has been positioned
for dull, dirty, and dangerous applications, tasks that humans don’t want to do.

One core measure that has made industrial robots so successful is that there is
a clear return on investment (ROI) and a measurable payback period for this
autonomous equipment. For any automation investment, the system must
return greater value than the cost of the solution.

The ROI for humanoid robots remains unclear. Humanoid robots will be able to
perform a variety of tasks, with each task having a different value to the robot
operator. Compare this to an industrial robot deployed into a specific task like
spray painting a car body or assembling a circuit board for an 8-hour shift - this
work is measurable and quantifiable.

6 of 95
The difficulty in determining the value of a humanoid robot comes when you
realize that a dynamically balancing humanoid robot is an order of magnitude
more complex than the state-of-the-art industrial systems or wheeled
autonomous mobile robots (AMRs) in production today. This makes it difficult to
size the market, and results in wildly varying market sizes and growth rates.

Several factors have sparked a “Cambrian explosion” of humanoid innovation in

the past few years. The cause is a unique nexus of factors:

• Computing power has reached a threshold and become small enough

that the computing power of the first Cray Supercomputers now fits
inside your smartphone.
• Battery technology has achieved a level of power density that can offer
reasonable runtimes for dynamic machines like humanoid robots.
• Artificial intelligence has evolved at an unprecedented rate over the past
five years.

Until now, the limits of computing, power, and AI have hindered the realization
of a humanoid robot form factor.

In researching this project, we collected data for over 160 different humanoid
robot models that are being developed around the globe by over 120
companies.[5] China and the Asia region in general lead the world with the
number of models and companies, and the Chinese government is pouring
billions of dollars into its domestic robotics industry. The U.S. and the Americas
are No. 2. Europe, the Middle East, and Africa (EMEA), and the rest of the world
(ROW) are in a distant third place. See Figure 1.

7 of 95
Figure 1 - Humanoid robot models by headquarter region, 2025. (n=169) Source:
ASTM Humanoid project database.

Figure 2 - Humanoid robot companies by headquarter region, 2025.[5] Source:

ASTM Humanoid project database.

8 of 95
There is currently no believable market-sizing estimate for humanoid robots.
However, all of the research consistently predicts a multibillion-dollar valuation
in the next five to 10 years. The most conservative and best informed is USD 2
billion by 2032 (Interact Analysis).[5]

Some projections include:

• USD 13.25 billion by 2029 (MarketsandMarkets),[1] USD 3.83 billion in the US

market alone by 2029 (MarketsandMarkets).[2]
• USD 76.97 billion by 2032 (SNS Insider).[3]
• USD 6.72 billion by 2034 (Cervicorn Consulting).[4]

Humanoid robots are unlikely to replace traditional automation systems, which

excel at the high-speed, high-precision, and heavy-lifting tasks that humanoids
currently cannot match. For a widespread shift to occur, the performance-to-
cost ratio of humanoids must surpass that of purpose-built machines, a
potential outcome driven by the scalability of their general-purpose nature.

Mobile robots have also evolved quickly in the past decade, and in the process,
many of the early-to-market companies have either been acquired or gone out
of business as these systems commoditize.

As stated earlier, autonomy deployed into industrial and logistics applications

requires provable ROIs. Industrial robot applications require that the robot
operate in a guarded work cell, separated from humans. Safety regulations and
protocols for these applications are mature and well-defined.

Collaborative robots can work near humans but must work to avoid contact. If
contact occurs, they must be force-limited to ensure humans are unlikely to be
hurt. Humanoids have similar characteristics to collaborative robots, with a key
difference: they can tip over and potentially harm a person nearby, even if not in
direct contact. Future humanoids will be developed to have the ability to touch
and hold humans, e.g., helping an elderly person out of bed. This is currently
beyond the collaborative standard.

The very nature of humanoid design is that these robots are likely to end up
sharing workspaces with humans. The majority of existing models have been
engineered to mimic the physical characteristics of humans. The average height

9 of 95
of the current crop of humanoids is 163 cm (64 in),[5] and the weight is 66 kg (145
lbs).[5] With two exceptions — 1X and Clone – the humanoid robots are all
covered in hard metal alloys, carbon fiber, or hard plastic.

Humanoid mobility is enabled by either bipedal walking or via a wheeled base.

Bipedal robots require dynamic stability, which means that there is always a
danger of the system tripping or falling to the ground. If there is a catastrophic
failure, the 60+ kg machine is likely to tumble or slump to the ground, trapping
anything underneath, including pets, toddlers, or the elderly.

Because of this inherent instability and catastrophic failure modes, humanoid

robots require an accepted safety standard for all machines before deployment
around the public or untrained persons.

Sources:

• https://www.marketsandmarkets.com/Market-Reports/humanoid-robot-
market-99567653.html
• https://www.marketsandmarkets.com/PressReleases/usa-humanoid-
robot.asp
• https://www.snsinsider.com/reports/humanoid-robot-market-1616
• https://www.cervicornconsulting.com/humanoid-robot-market
• ASTM humanoid robot database
• https://interactanalysis.com/insight/humanoid-robots-large-opportunity-
but-limited-uptake-in-the-short-to-mid-term/

10 of 95
 The Unique Risks Humanoids Bring and Why
Standards Must Evolve

Authors: Ingo Keller, Aaron Prather, and Rahul Ramakrishnan

Humanoid robots hold tremendous promise because of their human-centric

design. By mirroring our form and movement, they can utilize our tools,
navigate our spaces, and undertake tasks that once required human hands. This
very similarity, however, is what makes them uniquely complex and uniquely
risky.

Unlike industrial robots operating behind cages or mobile robots confined to

controlled routes, humanoids are meant to work with us, often in close physical
and social proximity. They blur the line between machine and partner, shifting
from simple cooperation (working alongside us on separate tasks) to deep
collaboration, where robots and humans share workflows, tools, and real-time
decision-making. This shift amplifies existing safety concerns and introduces
entirely new ones, from unpredictable physical behaviors to psychological and
ethical challenges.

The risks are multi-dimensional:

• Physical and Functional – Bipedal robots, by design, operate in states of

managed instability. A single loss of balance can have serious
consequences in shared spaces. Their dexterity still lags far behind the
human hand, creating hazards in tasks that require fine motor control.
• Psychosocial and Ethical – Their lifelike form encourages overtrust, leading
people to expect intelligence and emotional understanding far beyond
their true capability. In sensitive roles—such as healthcare, eldercare, and
childcare—this can erode trust, create emotional dependencies, or even
cause harm if performance falls short.
• Privacy and Cybersecurity – Constant sensing and networked connectivity
make humanoids potential surveillance tools and targets for malicious
control, raising concerns not just for safety, but for data security and
personal autonomy.

11 of 95
 • Reliability and Predictability – Unlike traditional robots, humanoids must
adapt to constantly changing human environments. A sensor glitch or
software fault isn’t just a technical failure—it can directly impact human
safety.

These challenges are not insurmountable, but they demand a new level of rigor.
Existing Robotics standards, designed for fixed, wheeled, or cooperative
systems, do not account for the dynamic balance, high-stakes collaboration,
and human-like interaction that humanoids bring. Simply extending current
safety requirements will not be enough.

How This Report Addresses These Challenges

The chapters that follow tackle these issues directly, offering a structured way
forward for manufacturers, researchers, and Standards Development
Organizations (SDOs):

• The first section examines classification, outlining why a shared

classification system is essential to defining humanoids, distinguishing
them from other robots, and mapping risks to specific applications.
• The second focuses on stability, identifying it as the most significant
technical and safety barrier to deployment, and proposing a roadmap for
performance metrics and safety validation tailored to actively balancing
robots.
• The third explores human-robot interaction, highlighting psychosocial,
ethical, and functional considerations that must guide how humanoids
are designed, tested, and introduced into shared spaces.

Together, these sections establish a framework for creating the standards

necessary to evolve humanoids from experimental prototypes into reliable,
certifiable tools. The specific risks mentioned above are not reasons to oppose
humanoids but are instead a call to address them differently through a
coordinated, evidence-based standards effort that reflects the complexity of the
technology.

12 of 95
 Bridging the Gap: From Risks to Standards

The risks humanoids introduce are no longer theoretical. As these robots move
out of controlled labs and into warehouses, hospitals, schools, and homes, the
challenges outlined above are already emerging in real deployments. The
question is not whether humanoids can perform the tasks, as they can
increasingly do so, but whether they can do so safely, predictably, and in ways
that humans will accept and trust.

Current standards only partially address this reality. Most were designed for
fixed or wheeled robots operating in either isolated industrial cells or highly
structured service roles. By contrast, they are generalists by design, capable of
working in environments that are not engineered for automation and
interacting directly with untrained users. This mismatch creates standards gaps
in three critical areas:

• Defining what kind of humanoid is being deployed (and for what level of
risk),
• Evaluating its stability and performance in dynamic, unstructured
settings, and
• Guiding human-robot interaction, physical and psychological, in diverse
populations.

Bridging these gaps requires more than simply adding new safety rules; it
demands insight tailored to the specific application. Not every humanoid use
case carries the same level of risk, and not every risk requires an entirely new
standard. Some scenarios can be managed with existing guidance, while others
represent critical barriers to deployment unless addressed through targeted
innovation, validation, or policy updates.

The following use case analysis examines how these risks manifest across key
sectors—manufacturing, healthcare, public services, and home environments—
and assesses where standards are sufficient, where they require adaptation, and
where entirely new frameworks may be necessary.

The table that follows provides this sector-by-sector risk view, setting the stage
for deeper discussions on how classification, stability, and human-robot
interaction standards can close these gaps.

13 of 95
R I S K R AT I O N A L E S FO R S E L E CT H U M A N O I D U S E CA S E

KEY STANDARDS GAPS /

USE CASE TOP RISKS
NEEDS

Extend ISO 10218 & ISO/TS

Physical safety in crowded
15066; ASTM ergonomic
WAREHOUSING OPERATIONS aisles
guidance for repetitive
Load handling ergonomics
material handling.

Physical safety near tools

Extend ISO 10218 & ISO/TS
Reliability in precision tasks
MANUFACTURING SUPPORT 15066; IEEE psychosocial and
Psychosocial impact on skilled
task-sharing guidelines.
labor

ISO 10218 & ISO 13482; UL/ISO

Reliability in handoffs
guidance for ergonomic tool
FACILITY MAINTENANCE Functional adaptability in
handoffs and adaptive
unstructured spaces
maintenance

Psychosocial overtrust ISO 13482, IEEE 7001; expand

CUSTOMER SERVICE & RE‐
Privacy & data use; ISO/IEC 24029 for AI trust; UL
CEPTION
Physical safety is manageable public HRI guidelines

IEEE 7010 & NIST IR 8269; new

Physical safety in crowds
SECURITY & MONITORING IEC frameworks for
Privacy & surveillance ethics
(PUBLIC) surveillance ethics & bias
Cybersecurity vulnerabilities
mitigation

ISO 13482 & NIST CSF; ASTM

Physical safety on roads
DELIVERY & LOGISTICS (OUT‐ protocols for environmental
Reliability in dynamic terrain
DOOR) adaptability; ISO road-
Cybersecurity threats
interaction safety

Privacy & cybersecurity

ISO 13482 & IEEE 7001; UL
GENERAL PURPOSE HOME Psychosocial overreliance
consumer certifications for
ASSISTANTS Functional limits in cluttered
residential robotics
homes

Physical safety in close care ISO 13482 & IEC 80601-2-77;

ELDERLY & DISABILITY SUP‐ Emotional dependence ISO care robotics standards;
PORT Privacy ASTM behavioral compliance
Ergonomic adaptability metrics

Psychosocial & ethical ISO 13482, ISO/IEC 29134, IEEE

CHILDCARE & EDUCATION sensitivity 7004; new ISO/IEEE
SUPPORT Privacy & data governance developmental safety &
Predictability in interaction ethical guidelines

14 of 95
The table highlights where humanoid deployment faces manageable versus
critical risks; however, context is crucial. Below is a brief overview of how these
risks appear across key use cases and what that means for standards
development.

Warehousing Operations

Humanoids in warehouses primarily handle picking, packing, and inventory

transport. Most risks are manageable, provided existing safety and functional
standards are applied. However, collaborative scenarios—such as handoffs or
navigation in crowded aisles—require additional ergonomic and stability
guidance to prevent collisions and improve load handling.

Standards Needs: Leverage ISO 10218, ISO/TS 15066, and IEC 61508; expand
ergonomic guidance under ISO or ASTM for repetitive material handling.

Manufacturing Support

Assembly lines and machine-tending roles require humanoids to work closely

with humans, where precision and timing are crucial. Physical safety, reliability,
and ergonomic limitations are significant barriers, compounded by the
psychosocial impacts of robots supplementing skilled labor. Fine motor control
remains a major limit.

Standards Needs: Extend ISO 10218 and ISO/TS 15066 for collaborative industrial
tasks; develop IEEE guidance for psychosocial impact and adaptive task-sharing
behaviors.

Facility Maintenance

Inspection and minor repair tasks generally carry moderate risks. Collaborative
handoffs—such as tool delivery—highlight the need for reliability and better
functional adaptability in unstructured environments, though psychosocial and
ethical risks remain low.

Standards Needs: Apply ISO 10218 and ISO 13482; create UL or ISO guidance for
adaptive maintenance behaviors and handoff ergonomics.

15 of 95
Customer Service & Reception

Public-facing indoor roles pose increased psychosocial and ethical risks due to
overtrust, unrealistic expectations, and concerns regarding data privacy.
Cybersecurity is essential to protect sensitive interactions, while physical risks
remain modest.

Standards Needs: Build on ISO 13482 and IEEE 7001; expand ISO/IEC 24029 for
public trust in decision-making and UL guidelines for public-facing HRI.

Security & Monitoring (Public Spaces)

Humanoids in security roles face some of the highest barriers. Operating in

crowded, unpredictable environments amplifies physical, ethical, and privacy
risks. Networked systems are particularly vulnerable to cyberattacks, raising the
stakes for both safety and trust.

Standards Needs: Integrate IEEE 7010 and NIST IR 8269; develop new IEC
frameworks for robotic surveillance ethics, bias mitigation, and active threat-
response reliability.

Delivery & Logistics (Outdoor)

Outdoor navigation introduces the most severe combination of risks—unstable

terrain, unpredictable human interaction, weather, and cyber threats.
Functional limitations and reliability issues are critical barriers to deployment.

Standards Needs: Expand ISO 13482 and NIST CSF; create ASTM protocols for
environmental adaptability and ISO standards for robot-road interaction safety.

General Purpose Home Assistants

Residential assistants face moderate but layered risks: privacy, cybersecurity,

and psychosocial impacts from overreliance or unmet expectations. Ergonomic
and functional limits are less critical but still require attention in cluttered home
environments.

Standards Needs: Strengthen ISO 13482 and IEEE 7001; develop UL consumer
safety certifications for residential humanoids.

16 of 95
Elderly & Disability Support

This is one of the highest-stakes use cases. Close physical assistance, health
monitoring, and emergency support demand rigorous safety, reliability, and
ethical oversight. Emotional dependence and privacy concerns are significant,
and ergonomic adaptability is often limited.

Standards Needs: Build on ISO 13482 and IEC 80601-2-77; introduce ISO
standards for human-centered care robotics and ASTM behavioral compliance
metrics.

Childcare & Education Support

Working with children raises heightened psychosocial and ethical concerns,

including data privacy and developmental impacts. Physical safety and
predictable behavior are non-negotiable, especially during play or collaborative
learning tasks.

Standards Needs: Extend ISO 13482, ISO/IEC 29134, and IEEE 7004; create new
ISO/IEEE standards for developmental appropriateness, safe interaction, and
responsible data handling.

17 of 95
 The Classification of Humanoid Robots

Authors: Benjamin Beiter, Brandon J. DeHart, William Harrison, Syrine

Mansour, and Thomas Mather

Introduction

One of the first recommendations regarding standards for humanoid robots is

to create a classification. Such a standard would classify humanoid robots into
different groups, and differentiate between individual humanoid robots and
between humanoids and other classes of robots. Safety is a primary concern for
both manufacturers and potential users of humanoid robots, and so it is a
primary factor that dictates the need for standards that provide the tools
needed to meet safety requirements. Classification is an important component
of standardization as well, as it could be used across all humanoid-related
standards, setting metrics, test methods, guides, certifications, and definitions
for specific robot classes.

A well-formed classification could be used to convey robot features, exclude

other features, identify additional relevant standards for a particular robot, as
well as understand design requirements, needed performance capabilities, and
safety restrictions for a specific class of robots. This chapter is not intended to be
a complete robot classification, but will discuss the challenges related to
classification, summarize prior standardized classifications of robots, and
highlight several potential approaches that could be used to build a complete
classification.

Defining a “Humanoid Robot” … Or Not?

The first question when considering the creation of standards for humanoid
robots is: “What is a humanoid robot?” While seemingly simple when
considered colloquially, there arise distinct discrepancies in any definition when
attempting to precisely define the categorical bounds of robots to which the
name applies. Does a robot need to have two arms and two legs to be a
humanoid? Does it need a head? Does that head need to have sensors just like
humans? How human-like does it need to look? What about size, weight, shape,
communication, mobility, strength, behaviors, identity………?

18 of 95
Innumerable characteristics could be used to define humanoids. This variance
in characteristics means that using any particular one in a definition would
result in overly restrictive categories that are not general enough to be useful to
manufacturers, users, or standards creators. The alternative is to define a
humanoid as a robot with any human-like characteristic at all, regardless of
application, functionality, or any other feature.

While not excluding any robots, this definition is also not useful, and would
require many subclassifications to sort humanoids into useful categories,
leading back into the first problem. As such, this chapter avoids defining
humanoid robots entirely, and instead recommends classifying robots by their
physical structure and capabilities, executive functionality, and use case/
application, just as is done in some current robotics standards.

This approach results in an overall classification system that can be used to sort
all robots, inclusive of humanoid robots. Relative to this report, in the same
effort to fully classify humanoid robots relative to other robotic applications, a
classification for all robots can be presented. This decoupling of the term
“humanoid” from classification will lead to a more useful classification of robots
by their function, capabilities, and use cases. Some of these classes may involve,
but will not wholly depend on, the robot’s appearance or anthropomorphic
structure, which are most often given as the qualitative defining factors of
humanoid robots..

Note: Any use of the term humanoid for the rest of this chapter will refer
broadly to any robot that someone may consider to be even partially
anthropomorphic.

Current Classifications

In general, the field of robotics is far too broad to facilitate a low-dimensional

classification, as robots are, at their core, an integration of many different types
of hardware and software that come together into an emergent system with an
almost infinite variety of possible designs, paired with an endless variety of
potential applications. Furthermore, because humanoid robots are general-
purpose robots, they are harder to classify compared to the more narrow scope
of classical robotic systems. This section will give examples of prior and current
robotic classification efforts, setting up a basis on which the classification of
humanoids could be built.

19 of 95
Prior Classification Efforts

Currently, there is no broadly accepted cohesive classification of robotic systems

by which to guide humanoid robot classification efforts in general. Furthermore,
there are only a few examples of industry-facing classification efforts with
robotics applications. ANSI R15.08 is effectively the only published standard with
any real classification coverage for robots, though it only covers mobile robots.
There have been various categorization, taxonomy, and ontology efforts
originating from academia and other places (Kirschner et al., 2025; Prestes et al.,
2013; Kunze et al., 2011; Kim et al., 2024); however, they often lack the industry
perspective.

R15.08

ANSI R15.08 provides a classification for industrial mobile manipulators, shown

in Figure 1. In general, humanoid robots that can move like a human can be
considered to be Type C Mobile Manipulators. Humanoids that do not move like
a human (are non-legged or are not mobile at all, just anthropomorphic) can
either be classified as industrial robots, if they have a manipulator, or none of
the above (“END” on the flowchart). However, this classification only focuses on
industrial robots, not any other application domain. It also only considers the
physical structure of the robot (i.e., possessing a manipulator arm or not), and
leaves out other relevant details such as the functional capability or
interactability of the robot. As such, additional classifications should be created
to deal with the uses of humanoid robots with varying capabilities, intended use
cases, and needed interaction capabilities.

20 of 95
Figure 1: Classification of Industrial Mobile Manipulators and applicable
standards from R15.08.

ASTM F48 (FORTHCOMING)

ASTM Committee F48 is currently working to create a classification for

exoskeletons. While not published yet, the effort to classify such a broad type of
robotic system is relevant to the effort with humanoid robots, as it also must
balance complex design, variable use cases, and performance measures (which
are hard to fully quantify).

SHANGHAI UNIVERSITY HUMANOID EMBODIED INTELLIGENCE STANDARD

(FORTHCOMING)

Finally, there is a forthcoming standard being developed as part of a national

initiative in China entitled: “Shanghai University Humanoid Embodied
Intelligence Standard”. This standard presents a classification of humanoid
robot ‘intelligence’, which is based on several different capability factors as
detailed in the standard.

21 of 95
Example of Possible Classifications

As can be seen from prior robotics classification approaches, there are varieties
of features upon which humanoids could be classified (a feature here being any
characteristic of the robot that can be used for differentiation). Prior
classifications are restricted in scope to specific applications or forms of robots.
Looking forward, however, humanoid robots are intended to be general-
purpose robots, and so can fill many roles and come in many forms, and as such,
most potential classifications that apply to any robots can and will apply to
humanoids as well.

The most prominent capabilities of humanoid robots (e.g., manipulation,

mobility, human interaction) are important features of almost all robots.
Therefore, it is recommended that instead of specifically creating a classification
of humanoid robots, effort be put into the general classification of all robots as a
necessary tool for both academics and industry to talk about specific kinds of
robots. Then humanoids can be classified by the same features as all robots to
best understand the use-cases, performance, and capabilities of individual
robots. An additional classification based on features unique to humanoids, or
at least common to humanoids but rare in other robots, may be useful, but
should fit within a larger robotic classification.

Based on the types of features that are often used to qualitatively describe
humanoids, there are many possible ways to classify them, and robots in
general. When creating classifications, the goal is to balance simplicity with
usefulness, often making a single scale of classes (e.g., Class 1 to Class 5).
However, a scalar approach inherently limits the generality and usefulness of
the classification. Without a coherent classification approach for robotics at the
moment, this chapter presents many possible features by which a classification
of robots could be made, focusing on features that also apply to humanoid
robots. The variety of features illustrates the complex considerations that go into
designing and implementing humanoid robots, as well as the challenge
presented by the prospect of creating a clear classification of robots.

In this section, we present a non-exhaustive list of example features by which

the classification of robots could be implemented in general. These features are
sorted into three categories, described in Table 1. This specific list of categories is
the most convenient to describe the features chosen here, but other categories
could readily be used (or added to this list).

22 of 95
TA B L E 1 : E XA M PL E S O F CAT E G O R I E S O F F E ATU R E S BY
WH I C H R O B OT S CA N B E C L A S S I F I E D.

APPLICATION: INTERACTIONS
PHYSICAL CAPABILITIES HIGH-LEVEL BEHAVIORS
AND RISK

FEATURES THAT ARE INHER‐

ENT TO THE DESIGN AND Ways the robot acts in certain
Tasks the robot is expected to
MOTION OF THE ROBOT. THIS situations, including with
perform, and in what
INCLUDES DYNAMIC MOTION safety-critical responses to
environment. These features
CAPABILITIES, SUCH AS LO‐ environmental inputs and
are not dependent on the
COMOTION AND MANIPULA‐ other observations. These
physical design of the robot
TION, AS WELL AS THE LOW- behaviors could be entirely
itself, but instead on what it
LEVEL MOTOR AND MOTION modified via the software and
will be expected to do.
CONTROL CAPABILITIES THE controls of the robot.
ROBOT POSSESSES.

Each of the classifications within these categories is defined by a set of integer

levels starting at 0 (indicating the system cannot perform the associated action)
and moving up, with higher values roughly translating to the increasing
complexity of controlling and using the physical system to successfully perform
the action associated with the design aspect. Classification levels may be
inclusive, so for example a level 3 robot could also have the capabilities from
levels 1 and 2, but would still be classified as level 3. It should be noted that
future classifications can make use of sub-levels as demonstrated by the E-stop
behavior feature. The different levels for each of these design aspects are
described in the following sections and are meant only as examples of possible
approaches to classifying robots; there are many viable alternative approaches
that a future standards committee could consider.

Physical Capability Features

These classification approaches are about the features that are inherent to the
design and motion of the robot. The features describe active capabilities that
the robot has, including features of its physical structure and design. The
features will change if the physical characteristics of the robot change,
especially for any robot that can transition between levels.

23 of 95
ENVIRONMENTAL MANIPULATION

This classification is based on what level of environmental manipulation the

robot is capable of. Specifically, it can be summarized as a measure of the
degrees of freedom (DOF) of the system’s manipulator(s), how many
manipulators it has, and whether cooperative manipulation is possible or not
with multiple manipulators.

Level 0: No environmental manipulation capability.

Level 1: Single manipulator with less than 3 DOF (e.g., shelf lifting attachment)
Level 2: Single manipulator with at least 3 DOF (e.g., typical industrial arm)
Level 3: Multiple 3+ DOF manipulators with fully independent workspaces
Level 4: Multiple 3+ DOF manipulators with partially shared workspaces

LOCOMOTION/MOBILITY

With locomotion, or mobility, it is difficult to define discrete transition points

between one level of capability and the next. For this classification, we will be
using the support polygon of the system, defined based on the locations and
types of contact points the system has with the environment. The support
polygon is a key indicator of static stability for any system being acted on by
gravity: provided the centre of mass of the system is “above” the support
polygon (in the direction opposite the force of gravity), then the system can be
maintained in static equilibrium using its current contact points with the
environment.

Level 0: Fixed-base robots, which are rigidly attached to their environment

Level 1: Mobile robots, which are statically stable (turning off leads to a stable
pose), as the support polygon is effectively constant and always under the
center of mass
Level 2: Balancing robots, where upright balance is only achieved with active
control due to the underactuated nature of their contact with the environment
Level 3: Stepping/Jumping robots, which are balancing robots that can
dynamically change their support polygon’s shape, size, and position by
moving one or more limbs
Level 4: Climbing/grasping robots, where the limbs used for mobility (legs,
arms, or otherwise) can make fully actuated environmental contacts (e.g., using
handrails)

24 of 95
HUMAN INTERACTION CAPABILITY

To evaluate the interaction capabilities of a system, without needing to also

incorporate an evaluation of the software and/or controllers generating any
related interactive behaviors, the focus must be on the physical attributes of the
system which are used to interact with people rather than how those physical
elements are used in any particular application or behavior. Therefore, the levels
defined for this category are predicated on evaluating what level of interaction
is possible given the design of a particular system without regard for whether or
not any interactive behaviors have been developed.

Level 0: No interaction capabilities

Level 1: No direct interaction capabilities, but capable of interacting with
people indirectly via a wired or wireless interface device (e.g., teach pendant)
Level 2: Includes lights, speakers, a screen, and/or other features to enable
basic one-way interaction with people in its local environment (e.g., turn
signals)
Level 3: Includes a camera, microphone, touchscreen, and/or other features to
enable two-way interaction with people in its local environment (e.g., verbal
communication)
Level 4: Includes an active head, human-friendly hands, and/or other features
to enable two-way physical interaction with people in its local environment

Behavior Features

Here, a behavior is defined as a series of coordinated actions to achieve a goal.

These types of behaviors are still dependent on the physical capabilities of the
robot, but are mostly related to and focused on how the robot acts in the real
world, based on the controllers, sensors, and other combinations of software
and hardware functionality and features that the robot has.

For this form of classification, categories can be defined related to how

dynamically the robot can adapt to human interaction, variations in payload,
adapting its locomotion based on sensing its environment, how intelligent it
appears to be to an observer, and even what its behavior might be when its e-
stop is pressed or other safety-related inputs are engaged.

25 of 95
DYNAMIC BEHAVIOR GENERATION

This outlines the level of interaction that exists for a person to affect how the
robot will perform its task(s) while the task, or behavior, is underway.

Level 0: Autonomous completion of pre-planned tasks, limited human-robot

interaction for task scripting only, before operation.
Level 1: Fixed workstation collaboration. Humans can continuously
communicate, command, and interact with the robot through an external UI.
Level 2: Robot can interact with humans directly (e.g., verbally, gestures) and
continuously through prescripted menus or other interface structures.
Level 3: Robot can interact with humans directly (e.g, verbally, gestures) and
continuously, not requiring any user interface. (human-equivalent interaction)

DYNAMIC PAYLOAD COMPENSATION

How capable the robot is at adjusting its planned motions based on the payload
that it is carrying. Here, the term “Hard-coded” refers to pre-identified, assumed
inertial parameters of the payload, while “Adaptive” refers to online
identification and/or compensation of measured inertial parameters of the
payload.

Level 0: No dynamic compensation

Level 1: Hard-coded gravity compensation of known payloads
Level 2: Hard-coded gravity and inertial compensation of known payloads
Level 3: Adaptive gravity compensation of measured payload parameters
Level 4: Adaptive gravity and inertial compensation of measured payload
parameters

APPLIED LEARNING

This captures the extent models learned from data determine the behaviors of
the robot.

Level 0: Basic autonomy, following pre-programmed controls and discrete

logic for behavior generation. No learning.
Level 1: Models have been trained and employed for specific uses/processing,
but overall behavior generation is still explicitly coded.

26 of 95
Level 2: Underlying behaviors are still traditionally controlled, but high-level
action planning is done by a pre-learned policy.
Level 3: Executive action planning is done by a policy that continues to learn
in operation.
Level 4: End-to-end learning, or at least all processes being effected and
governed by an integrated, continuously learning policy.

LOCOMOTION PLANNING

This capability captures the extent to which a robot can plan and then follow a
path plan in a variety of environments. More complex environments require
more complex sensing, planning, and motion control algorithms.

Level 0: Can follow a pre-planned set of footsteps.

Level 1: Periodically replan footsteps towards locomotion goal on simple
terrain (flat, no obstacles)
Level 2: Can replan footsteps based on static, complex terrain (knowledge of
terrain can come from onboard vision/mapping or an oracle)
Level 3: Can replan footsteps based on changing, complex terrain, including
added obstacles, fragile ground / infeasible footstep locations, or physical
disturbances.
Level 4: Can plan whole-body environment interactions, utilizing both hand
and foot interactions for optimal plans.

E-STOP, FAULT, OR POWER-LOSS BEHAVIOR

The fault-response behavior of the robot. Risk cannot be completely removed,

but more mitigation factors can reduce the risk posed to nearby users and the
environment. An individual robot would be classified based on the highest level
of behavior it is capable of, even though it may be able to demonstrate lower-
level behaviors.

Level 0: Immediate power cut to all components.

Level 1: Brakes (either powered or unpowered) are employed to enforce
passivity/dissipation of total energy in the system.
Level 2: Robot actively reduces its footprint / size / effective endangered area
by curling up, leading it to cover pressure points and other dangerous surfaces
on the robot. (Action is irrespective of surroundings)

27 of 95
Level 3: Robot views surroundings, then plans and takes action to quickly and
as safely as possible move to a safe robot state. This specifically addresses the
hazard, e-stop, or other fault that initiated the hazard-response state.
Level 3-1: Robot additionally maintains power enough in joints, even after a
main power source loss, to enact these safety actions.
Level 4: Robot is always operating in a way such that failure of components or
power loss will not endanger nearby humans or other designated protected
objects. (e.g., maintain a pose at all times such that a power-loss collapse will
occur away from a human)

Application: Interactions and Risk Features

An interaction and risk-based classification addresses many of the key safety

concerns surrounding the environments in which a humanoid robot may exist.
Robots' interactions in various types of environments induce different levels of
risk. With the following features, risks are described qualitatively with a
comparative scale based on how the robot is expected to interact with the
environment.

When a robot does not have to interact with humans in an environment, it can
act autonomously. Multiple robots can communicate individually or with a
centralized control. When in a mixed environment, meaning both humans and
robots are present, the risk depends not only on the robots’ control but also on
the background of the present humans. Having only trained humans in the
environment means they know how to behave in a shared space, are aware of
the instructions, and are adept at handling different scenarios.

Mediated human surroundings refer to a controlled human environment with a

mix of trained and untrained participants; the trained participants would take
the lead or supervise whenever an interactive situation occurs. A public
environment is completely uncontrolled, where participants do not take specific
actions and may be unaware of how to respond across a diverse range of
interactions. The type of environment, and thus expected interactions, is also
dependent on the application the robot is being used for.

INTERACTION COMPLEXITY

Refers to the variety of other actors that the robot must be able to interact with,
including both other robots and humans of various amounts of experience.

28 of 95
Level 0: Fully Separated - inline with current industrial standards (e.g.,
traditional industrial robots in fenced zones with light curtains, safety PLCs
triggering shutdowns). Interaction is prevented.
Level 1: Interaction with other robots (e.g., fleets of AMRs in a warehouse).
Level 2: Interaction with trained humans (e.g., collaborative robots on
assembly lines working alongside knowledgeable staff, potentially involving
remote operation).
Level 3: Interaction with untrained humans (e.g., service robots in public
spaces, delivery robots on sidewalks, potential future home assistants).
Level 4: Interaction with Vulnerable Populations (children, seniors, medical
patient, etc.)

INTERACTION TYPE

Refers to the mode of interaction that the robot must be capable of, from digital
communication with a central server, to physical communication and
interaction with humans.

Level 0: Interact with no other robots or humans

Level 1: No direct interaction, but communicates and receives commands
from a central fleet manager system.
Level 2: Direct Communication with other robots
Level 3: Direct Physical Interaction with other robots
Level 4: Direct Communication with humans
Level 5: Direct Physical Interaction with humans

APPLICATION DOMAIN

This refers to the general application the robot will be used for. It implies the
environment, interactions, expected actions, and capabilities the robot must
have, as well as the most common disturbances, safety considerations, and
design requirements that are expected from a robot in this application domain.
This classification breaks with the numeric, performance-based metric
structure.

• Industrial - The robot is used for job-related applications, often in

structured environments with set tasks.
• Residential - The robot is for personal use in homes. The environment is
unstructured and/or changing, but the tasks may remain constant.

29 of 95
 • Service - The robot is implemented and used by an owner to interface and
interact with the public.
• Public - The robot must autonomously act in an unstructured
environment, and adjust plans and actions based on observations of the
surroundings.
…..

Note: Many other application domains could be listed, such as construction,

medical, underwater, etc. The above is only a sample of possible application
domains.

Classification of Humanoids

Humanoid robots, or rather “robots with human features”, are not necessarially
novel in their capabilities or features, as many other robotic systems can share
those details.. Humanoids do, however, indicate a clear expansion of the types of
environments robots will be deployed in. Deployable applications will move
from isolated cages into shared spaces, necessitating layers of new safety and
performance requirements. Currently, robots existing in shared spaces are
typically either small or purely animatronic. Robots capable of the speed and
strength of industrial robots are usually completely isolated in cages or at least
only deployed in industrial environments around trained personnel. Humanoid
robots will potentially have the characteristics of today’s industrial robots while
also existing in shared spaces around untrained people. A classification of
humanoids could thus be used to understand the safety expectations and
performance requirements associated with the robot and its behaviors in these
environments. Consumers want to know what to expect when products come
to market, and manufacturers want to know what kind of capabilities and
functions need to be built into their robots.

Even though humanoid robots do share characteristics with other robotic

systems, it can still be useful to have a classification of humanoid robots, based
on features common to all humanoids, or features that are unique to
humanoids. One such unique feature is humanoid robots’ humanlike
appearance. A classification based on visual features is hard to define, especially
when those features may not provide any function or have questionable use.
However, one potential method of classification, as outlined below, is based on
the purpose of the human form, which could be used to sort the different
humanoid robots that are currently being designed and manufactured. It is

30 of 95
useful for capturing the intent behind the creation of the robot, describing why
the robot that is used for an application should be a humanoid rather than a
more classical robot structure. Note again that this is an example of a
classification approach for humanoid robots, and is meant only as a starting
point from which future classification efforts can begin.

REASON FOR HUMAN FORM

For what reason has this robot been made to look like a human?

Class 0: Has no human form

Class 1: Purely visual appeal (appearance does not convey the expected
function or capability of the robot)
Class 2: Leveraging form to affect the reaction (for example: physical,
emotional, psychological, etc.) users have to the robot
Class 3: Acting in environments built for humans
Class 4: Replicating human capability for general action

Conclusion

The effort to develop a classification for humanoids uncovers the larger need for
a robotic classification in general. The classifications presented above are
example strategies for classifying all robots. However, because the classifications
are based on features that humanoid robots also possess, the same
classifications can be used to sort humanoids by capabilities and features. Once
a classification for all robots has been created, then a classification for
humanoids can be created within that context, adding features that are unique
to humanoids, such as the similarity to human form, as discussed above.
Furthermore, in the interest of engaging a broad range of participants in the
standards creation process, if any readers had a strong reaction, opinion,
critique, or idea about these example classifications, they should join the on
follow-up committee that is creating a standard on robot classification after this
report.

Chapter Bibliography

• Kirschner, Robin Jeanne, et al. "Categorizing robots by performance fitness

into the tree of robots." Nature Machine Intelligence (2025): 1-12.

31 of 95
• Prestes, Edson, et al. "Towards a core ontology for robotics and
automation." Robotics and Autonomous Systems 61.11 (2013): 1193-1204.
• Kunze, Lars, Tobias Roehm, and Michael Beetz. "Towards semantic robot
description languages." 2011 IEEE International Conference on Robotics
and Automation. IEEE, 2011.
• Kim, Stephanie, Jacy Reese Anthis, and Sarah Sebo. "A taxonomy of robot
autonomy for human-robot interaction." Proceedings of the 2024 ACM/
IEEE International Conference on Human-Robot Interaction. 2024.

32 of 95
 The Role of Stability

Authors: Ben Beiter, Brandon J. DeHart, Federico Vicentini, Francisco

Andrade, and Dieter Volpert

For stability, in this chapter, we are considering classic humanoid robots that are
bipedal, powered, and actively-balancing. Such robots differ from other types of
robots for which we have standards by the fact that they can only stand
through powered balance and can change the shape and position of their
support polygon relative to the rest of their structure during normal operation
(i.e., they can take steps). These mobility features provide the agility, range of
possible motions, and responsiveness that give humanoids the highly varied
potential use cases they have. However, these same capabilities are also the
main source of hazards caused by humanoid-type robots. At all times during
operation, a loss of power or too large an unexpected disturbance could result in
the robot toppling over. This could harm the robot itself, any parts of its
environment, and any nearby people. As a result of this risk, stability-related
safety concerns are a main barrier to the adoption of humanoids in any space
shared with humans.

The proper way to address these safety concerns would be through a risk
assessment that identifies all potential hazards and the appropriate safety
measures that must be taken in response to those hazards. Ideally, technology-
and application-specific standards would be available to guide readers
(standards developers, policy makers, robot researchers, etc.) through the risk
assessment process. Currently, however, there is a lack of standards that provide
the tools needed to quantify the level of risk or validate the effectiveness of
safety functions on actively-balancing robots. The entire burden of proving that
the robots are safe enough is thus on the manufacturer of each robot. Creating
standards for evaluating stability will ease this burden on manufacturers. Such
standards will be used to measure and prove stability during key applications,
allowing customers to have confidence in the safety of the robot, as well as
allowing manufacturers to measure performance on tasks and prove utility for
the applications that customers want to use them for. Creating practically
useful standards is the necessary first step towards creating certifications for
humanoid robots.

33 of 95
An example first step is to address the fact that standards currently do not have
a cohesive definition of stability. The word itself means slightly different things
in different contexts. The technical engineering definition of a system state
settling to a defined equilibrium, the colloquial sense of a system able to act
and interact with the environment without harming anything, and the specific
humanoid sense of being able to walk and balance without falling over, or at
least managing the risk of toppling in the same way humans do. In addition to
the varying types of stability, there are two aspects of stability to be considered:
safety and performance. Safety considers the avoidance of harm to people in
the vicinity. This can be considered a minimum requirement and more of a
pass/fail behavior. Performance considers evaluating capabilities to both
perform intended tasks and respond to changing conditions, which can also
include measuring potential risks posed to the robot itself and the environment
as a desirable byproduct.

Despite not being inherently stable, humanoid robots are expected to be able
to perform a wide variety of tasks while also being robust to a wide variety of
environments and external disturbances. On the other hand, the complex
structure and walking capability of these robots give them many options in how
to respond to events and disturbances (e.g., navigating variable terrains or
stepping to avoid falling rather than just leaning to balance). Current robots
have displayed increasingly impressive capabilities as they approach
applications that live up to the high expectations of humanoids. However, the
variation in how a robot might respond to certain situations, as well as the
variety of control approaches that drive such behaviors, makes evaluating the
stability of humanoid robots more challenging. Additionally, because the
functioning of the robot is so dependent on the performance of the control
algorithm, if a specific combination of task, robot, environment, and inputs has
not been seen and tested before, any standardized validation space would be
very broad, at the limits of practicability.

There is a need for metrics, test methods, design procedures, and

implementation guidelines for humanoid robots that can guide the use of
humanoid robots in workplaces. This chapter outlines a recommended multi-
layered approach for creating standards that help quantify the high-performing
potential applications of humanoids and provide requirements to verify the
safeguards implemented to protect humans, equipment, and the environment
in which the robot is deployed. Section 1 reviews current standards and prior
research related to humanoids. Section 2 outlines the recommended roadmap

34 of 95
of stability-related standards that will be detailed. Within that roadmap, Section
3 discusses the need for quantifiable performance metrics and methods to aid
in understanding humanoid robot motions, and Section 4 discusses how the
stability standards are closely related to the development of safety standards as
well. Section 5 provides concrete recommendations for Standards Development
Organizations (SDOs), and Section 6 concludes the chapter.

Review of Current Safety Standards

For robotics, considerations of safety begin with safety standards for machinery.
Safety for machinery is implemented in two steps: risk assessment and risk
mitigation. Risk assessment for machines is guided by ISO 12100, with other
standards covering individual classes of risks, such as those derived from
electrical hazards (IEC 60204-1). The fundamental hazards for machinery in ISO
12100 are extended in product-specific safety standards like industrial robots
(ISO 10218-1 and -2, recently revised), personal care robots (ISO 13482), and
others. Risk mitigation is often done in one of two ways: either avoiding a
particular risk entirely through safety requirements or actively avoiding it during
operation through safety functions. Some safety requirements inside the
product-specific safety standards point to technology-specific or narrow
standards, for instance, the ISO 13850 on Emergency Stop or the ISO 13855 on
safe distances from static and moving hazards. Product-specific standards
address vertically the risk assessment for robots, the selection of equipment,
both electrical and mechanical, the properties of controls and power sources,
the mandatory safety-related stops, mechanical and non-mechanical physical
or virtual safety guarding to protect people, motion limits such as avoiding
kinematic singularities or speed and separation monitoring, labelling,
instructions, and more. When these protective measures are implemented with
devices dedicated to the reduction of risks via the monitoring and active control
of reaching or maintaining safe states, then Functional Safety is implemented
following IEC 61508 (or its derivation for machinery ISO 13849). Altogether, these
standards establish general requirements for the safety of machinery, and
robots specifically.

Beyond these general safety standards, ISO and other SDOs have produced
several safety standards that address more subcategories of robots. However,
these standards make an unwritten assumption that the base of the robots
being considered is either fixed or has a statically stable base. Due to this, none

35 of 95
of these specialized safety standards apply to humanoids. For humanoids, even
with all the joints braked and powered off, which is a very conventional safe
state for fixed manipulators, the robot is not necessarily in a safe state. In
general, it must be assumed that a locked pose is unstable. This potential
instability, present in normal conditions, means that the robot can become
unbalanced and fall over even with high-performing controllers. Current
standards do not adequately address stability requirements, and so a dedicated
series of standards is now recognized as a market need.

The only standard that does not exclude robots with actively controlled stability,
including legged robots, from its scope is ANSI/RIA R15.08-1. This US national
standard provides safety standards for industrial mobile robots (IMRs), including
their navigation and control. R15.08-1 is uniquely silent concerning the mobility
principle of robots, and legged robots with manipulators could fit in the scope.
The standard defines “Type C IMRs” when manipulators are integrated with a
generic mobile base, which is the mobility representation closest to humanoids.
While R15.08-1 states that “The IMR shall maintain stable operation during
travel” and includes some methods to validate this stability condition, it only
considers that instability could be caused by a payload that is too heavy and/or
held too far away from the base of the robot, causing the IMR to tip over. These
requirements are derived from the risks analyzed for single manipulators
mounted on top of moving bases.

This limited approach to stability is mirrored in other available safety standards.

The ANSI/ITSDF B56.5 for industrial trucks (e.g., AGVs) includes dynamic stability
considerations, but they only account for the inertia of the vehicle and do not
cover any balancing robots with a changeable base of support. B56.5 also states
that “The user shall be responsible for the load stability and retention. When
deemed necessary by the user, verification shall be required,” though it does not
present any test methods or requirements for controlled stability.

For manipulation, no safety standards directly address the effects of

manipulation on stability. All requirements for robot arms are in ISO 10218, while
ISO/TR 20218-1 further illustrates hazards for end effectors and their interfaces,
stating, for instance, that robots should ensure “that loss of power does not lead
to loss of load, unexpected motion, or other hazards”. This might be
straightforward for fixed-base manipulator robot arms or whenever the reaction
energy is completely absorbed or balanced by a stable base, but humanoids do
not follow this rule, as losing power at any time could result in falling over.

36 of 95
Humanoids are far more complex in the aspects of dynamic stability and the
use of limbs for and during balance. A dedicated family of standards that
address stability for actively-balancing and legged robots would greatly
contribute to the clarity of specifications and requirements about manipulation
and balance.

In the domain of Service Robotics, IEC 63310 on Active Assisted Living Robots
states, “AAL robots with assisted mobility functions other than wearable or
wheeled ones should have the ability to finish their intended tasks.” ISO 13482
on Personal Care Robots simply states that robots should have sufficient
stability: “The personal care robot shall be designed to minimize mechanical
instability (e.g. overturning, falling, or excessive leaning when in motion) due to
failure or reasonable foreseeable misuse.” It references ISO 7176-1 and ISO 7176-2,
which are about static and dynamic stability for wheelchairs. Again, these
standards only determine that the device should not be destabilized while in
normal use. It does not address robots with dynamic self-balancing or the ability
to alter their base of support during operation.

Even further from industrial and service robotics safety standards, a potential
use of humanoid robots appears in the Technical Report ISO 4448-1 about
Public-area Mobile Robots (PMRs). This report is form-agnostic and explicitly
includes legged robots, as well as the more common wheeled mobile robots,
and includes safety considerations for the application of automated pick-up and
drop-off (PUDO) of goods and people at the interface between roadways and
sidewalks (the “kerb”), in addition to the behavioral requirements for PMRs
operating in pedestrian spaces, including sidewalks and roadways. The core
goal of this report is to provide a framework for municipal, provincial/state, and
federal governments to establish regulations surrounding the operation and
use of automated systems performing PUDO tasks. Interestingly, the report
introduces the use of a “shy distance” which dictates how far away a PMR
should attempt to stay at a minimum from any “inattentive, uninvolved,
unprotected, and untrained bystanders”. For comparison, all existing mobile
robot standards do not dictate any predetermined separation distance, but refer
to a more sophisticated computation of sufficient space to come to a safe stop
or prevent hazardous collisions. Additionally, the problems of exposure of the
general population and outdoor conditions are very challenging.

37 of 95
Overall, most standards require that robots, whether mobile manipulators or
personal care robots, must not cause undue hazards due to instability during
normal operations, nor should maintaining stability increase the risks of already
existing hazards. First and foremost, “preventing hazards” due to loss of stability
is not an option for actively controlled robots. By the very nature of actively
controlled stability, hazards are not eliminated. Only risk reduction and the
evaluation of residual risks of instability are eligible considerations. Second, and
challenging from both technical and regulatory standpoints, balance-related
actions are part of the safety system in the scope of functional safety. A safety
function is the capability to detect a particular hazardous effect (i.e., a quantity)
through sensing of internal states (e.g., a fault) or external phenomena (e.g., an
obstacle), then enact corrective actions that successfully either avoid any
hazards or reduce the consequences of the hazardous event to acceptable
levels. Conventionally, the primary approaches to establishing confidence that
safety-related control actions are sufficient to reduce risks are by measuring
their ability to control or avoid failures according to standardized levels of
residual failures. This metric is established in IEC 61508 with Safety Integrity
Levels (SIL), and inherited with variants by the derived sector-specific functional
safety standards like ISO 13849 for machinery with Performance Levels (PL).
Both SIL and PL measures are based on intervals of Probability of Dangerous
Failures per Hour that are suitable for random failures and are paired to a
corresponding level of control or avoidance of systematic failures (Systematic
Capability). These levels are used to create requirements for designing
increasingly robust safety functions intended to reduce risks that are
proportional to the frequency and intensity of the associated hazards. The
higher the risk, the more demanding the protective function must be.

However, these functional safety standards are progressively not fully adequate
when hardware and software failures are intertwined. This is the case of actively
controlled stability and perception of the surrounding environment that is
instrumental in achieving successful locomotion and navigation.

For example, typical top-risk safety functions for industrial robots are mandated
in ISO 10218 to achieve a PL d range of residual dangerous failures, specifying
the required range of those probabilities depending on the hardware
architecture (structure category 2 or 3), when expressed per ISO 13849 - or to
reach SIL 2 with a hardware fault tolerance of 1, when expressed in accordance
to IEC 61508 or IEC 62061. This approach is deeply understood and consolidated
for relatively simple functions. However, the dynamically balanced nature of

38 of 95
humanoid motion means that any safety function related to stability will have
varying results and a chance of success depending on both the state of the
robot and any external factors that might be present. Recovering, or just
maintaining balance, requires a complicated series of sensing, internal
modelling, motion planning, and control capabilities that are harder to
characterize as SIL or PL safety performance levels. Additionally, there are
currently no test methods or procedures to measure performance and validate
the capability of balance safety functions. To have full confidence in the
performance of stability safety functions, more than just the standard SIL and
PL process is needed.

Hazards related to instability are, in fact, not well expressed in terms of rates of
failures per hour, but rather as events that depend on the execution of actual
behaviors and the conditions of the environment. Hardware and software
components implementing safety functions can be characterized by
distributions of failures with time-based rates, but aggregated behaviors are
better expressed in terms of the rate of success, and risks may still arise even
without system failures. This scenario is common in the domain of autonomous
vehicles, where methodology for the analysis of functional insufficiencies is a
preliminary step to instruct safety functions to address the ultimate execution
of a safety action (see ISO 21448, in combination with the functional safety in
ISO 26262). Similarly, assessing the safety of complex systems like humanoid
robots requires a comparable evolution in approach, focusing not just on
preventing/detecting malfunctions, but on ensuring an acceptable level of
safety of their intended actions within unpredictable environments.

Finally, an important consideration for dynamically stable robots is failure

handling, even during the execution of a safety function. What does the robot
do when something goes wrong? Most safety functions measure success by the
capability of the system to return to a safe state. Most safety standards have the
phrasing that “No new hazards shall be introduced during a stop”.

However, if it is already moving, a humanoid can require additional motion

planning and/or intermediary actions to reach a stable, safe state. These
intermediary actions (e.g., recoiling, balance compensation by parts of the body)
could add new hazards or concerns, even if they help the robot avoid falling
over. Guidance from safety standards for application-specific environments will
be needed to determine the proper behavior in certain situations, and thus be
able to properly consider humanoids in risk assessments.

39 of 95
In conclusion, current standards created requirements for safety with the
underlying application domains and their risks in mind (collaborative
applications, service robots, AGVs, AMRs, autonomous vehicles, etc.). Those
standards cannot be applied directly to humanoids, where the problem of
stability affects all safety-critical functions and the degree of successful
resolution to safe states. Dedicated risk assessment considerations and analysis
of functional insufficiencies are not common in the domain of robotics but are
concluded to be necessary for complex dynamic behaviors and highly variable
autonomous functions. On top of this, even the ability of humanoids to attain
balance and/or recovery implemented as safety-related parts of control systems,
there is a lack of standardized test methods to validate such safety capabilities.

All of this should be addressed early in the standards creation timeline.

Currently, it is part of the mission for the newly created working group in the
ISO Technical Committee 299 tasked to develop a safety standard (ISO/AWI
25785-1) for industrial mobile robots with actively controlled stability, notably
including humanoids.

Review of Current Test Methods and Performance Standards

Beyond safety, there is a need to be able to measure the performance of a

humanoid. Safety is the avoidance of hazards, and especially doing harm to
humans, while measuring performance is how capability and productivity can
be quantified and validated. Standardized metrics and test methods are the
tools to measure performance. While there are some ongoing efforts to create
performance test methods for humanoid and similar systems, such as through
ASTM International Subcommittee F45.06 on Legged Robot Systems, there are,
so far, no published standard test methods that consider humanoids
specifically. However, because humanoids are general-purpose robots, meant to
perform a wide variety of tasks, current task and application-specific standards
could be used to measure certain aspects of a humanoid’s capabilities,
although users of standards would need to make an extra effort to interpret and
adapt them.

For example, ISO 9283 focuses on establishing the accuracy and repeatability of
industrial robots applied to manipulation tasks. These tests are mostly applied
to fixed-base robot arms, with accuracies on the order of millimeters.
Humanoids can also perform manipulation tasks; however, this particular
standard is limited in that it only considers open-loop performance, not

40 of 95
covering the integration of the large number of sensing systems that
humanoids have. It also only focuses on end-effector positioning, not
addressing robots without a fixed base. This makes the test methods not fully
applicable as is, because unmeasured error in the positioning of the base will
interfere with manipulator performance. Similarly, ISO 18646 is a series of
performance measurements and test methods for service robots. The tests do
not consider stability beyond static loads offset from the COM or inertia from
dynamic motions, which cause tip over, with relevant work undertaken in the
ISO/DIS 18646-5 project (which is in final development stages at the date of
writing). Overall, for all robotics test methods that only consider fixed-base
robots, while the method itself may be able to be accomplished by the
humanoid, balance and stability performance will affect the overall results of all
of those test methods and must be considered.

Another source of applicable performance standards is the set of test methods

created for other systems that are not humanoids, but have some common
aspects in mobility-agnostic application scenarios. For example, the ASTM
International Subcommittee E54.09 on Response Robots published 20+ test
methods that define many apparatuses, tasks, and procedures that do not
depend on locomotion principles or capabilities. Other legged robots (primarily
quadrupeds) are regularly tested with these methods, which cover many
elemental behaviors included in navigation, coarse manipulation, and
inspection. Many of these methods, with some possible alterations for scale and
difficulty, would be directly applicable to humanoids, some even being useful
for testing stability and balance. For most standard test methods considering
mobile-base robots, there is nothing keeping them from being used to evaluate
humanoids as well; there is just a need for more and more targeted methods
and metrics to evaluate the specific template motions that a humanoid
employs. Metrics such as these will need to evaluate the performance of both
humanoids as a whole and their individual subsystems and capabilities.

As inspiration for new standard test methods, there have also been several
academic projects dedicated to evaluating humanoid robots in the past. There
are many papers that discuss humanoid robot benchmarking, though most
focus on finite aspects of humanoid robots such as WHole-Body-Control
approaches, learned control policies, specific tasks, or individual robots (as a
non-exhaustive list). Of note is the EUROBench Project, which was a 5-year,
multi-institution project that set out to create a unified benchmarking
framework for robotic systems in Europe. Various sub-projects generated test

41 of 95
methods for many aspects of robotics applicable to humanoids, from whole-
body manipulation to balance. Of all the research mentioned here, however,
none has yet to make it into any standards as metrics, test methods, or
practices. These projects are starting points for future committees to begin
forming standards for evaluating human performance.

Stability Standards Roadmap

To address the challenges that stability presents for the adoption of humanoid
technology, we recommend an approach that:

• Keeps accruing experience in the use of humanoid robots in the presence

of humans, without entirely blocking the adoption of technology due to
the lack of dedicated standards.
• Coordinates and fosters the definition of practical metrics, test methods,
and validation conditions to fully evaluate humanoid stability. This
element is very important for the quantifiable and objective evaluation of
stability and for establishing acceptable limits for safety.
• Creates dedicated safety standards. Safety standards have a deep
influence on regulatory matters, so clarity and limited room for
interpretation would be beneficial.

To establish both measurement and safety requirements, the to-be-proposed

standards may share key information currently existing for other robotics
applications. In particular, performance standards elaborating testing criteria
and methods with efforts undertaken by various SDOs. This will allow the
definition of minimum stability performance thresholds. The same metrics will
allow for potential live monitoring of stability, with the capability to enter safety-
critical control modes if certain conditions are met.

Path Forward

Across all prior research and standards, we have found that despite balance/
stability control being done on all humanoid robots, balance/stability
measurement has not been published in any standards yet, though several
programs for standard development are ongoing. Given the body of knowledge,
however, stability criteria should become part of accessible standards with a
high potential for benefit to the humanoid industry. Despite this clear goal,
implementing a codified set of criteria in technical standards is still not simple.

42 of 95
Therefore, we recommend a two-part process for creating new standards
focused on humanoid stability performance. The first is on measuring and
quantifying stability. This first effort is focused on creating test methods to test
stability performance in a variety of tasks and conditions. These tests will
become the foundation for overall humanoid performance testing in the future.
This effort will also focus on creating usable stability metrics. The second part is
on developing safety standards for humanoids. This will involve building upon
the current approaches to robot safety to extend to the particular hazards
presented by humanoids. The design of safety standards will come from two
directions. One will build on the test method and metric creation, where each
test method can establish safety thresholds for each task, and the other will be a
new approach to standards, looking at integrating safety requirements into the
controller itself at the design stage.

TA B L E 1 : A S U M M A RY O F CO N T R I B UT I O N S N E E D E D I N
N EW STA N DA R D S , A S WE L L A S R E CO M M E N DAT I O N S
F O R A S S O C I AT E D N E XT ST E P S TO D EVE LO P T H O S E
STA N DA R D S

NEEDED STANDARDS RECOMMENDATIONS

Consolidate metrics and test methods to

METRICS AND TEST METHODS THAT CAN evaluate stability performance
VERIFY REQUIRED PERFORMANCE CAPABIL‐ Identify the subset of metrics and stability
ITIES WHEN A ROBOT ENTERS AN UNSAFE tests that are safety-critical
STATE. Create specific validation configurations and
scenarios

Consolidate metrics and test methods to

evaluate stability performance
AGREED UPON STABILITY METRICS FOR ON‐
Identify the subset of metrics and stability
LINE MEASUREMENT OR ESTIMATION OF
tests that are safety-critical
SAFETY-CRITICAL STABILITY
Create specific validation configurations and
scenarios

Discuss safety standards for potentially

unstable mobile robots
Establishment of minimum thresholds of task
AGREED UPON CONDITIONS FOR SAFETY- performance required for safety, depending
RELEVANT LOSS OF STABILITY on environmental and application hazards
Online measurement of stability that a robot
can react to, depending on the safety
requirements

43 of 95
 NEEDED STANDARDS RECOMMENDATIONS

Discuss safety standards for potentially

unstable mobile robots
RISK EVALUATION OR SAFETY TARGETS FOR Establishment of minimum thresholds of task
THE IMPLEMENTATION OF ROBOTS, SUB‐ performance required for safety, depending
JECT TO INSTABILITY IN A SPACE THAT on environmental and application hazards
HUMANS MIGHT ALSO OCCUPY Online measurement of stability that a robot
can react to, depending on the safety
requirements

Discuss safety standards for potentially

unstable mobile robots
ACCEPTABLE BEHAVIORAL REQUIREMENTS Establishment of minimum thresholds of task
TO AVOID SITUATIONS THAT COULD LEAD TO performance required for safety, depending
HARM TO HUMANS FROM A ROBOT on environmental and application hazards
TOPPLING OVER. Online measurement of stability that a robot
can react to, depending on the safety
requirements

Quantifiable Performances: Test Methods and Metrics

A key need to be addressed for humanoid adoption is the capability to measure

and validate stability performance, which captures the stability of the robot
while interacting with any environment. This scope is broader than human
safety. We need to consider the risk of losing stability posed to the robot itself
and other surrounding equipment within the environment. Like safety for
humans, we cannot practically guarantee no risk to the environment from an
underactuated robot moving through it, but demonstrated performance on a
set of representative tasks can establish confidence that future performance will
be both safe and high-performing enough. To quantify this performance, there
are two necessary parts: (i) test methods and guides for global attributes that a
robot is to be evaluated on; (ii) objective, observation-based metrics to score the
stability of a robot during specific tasks.

The goal of these performance standards is to provide a way for manufacturers,

customers, and regulators to understand, evaluate, and compare the
performance of robots and the subsystems that make them up. Customers wish
to purchase robots that meet their application requirements, manufacturers
wish to shape designs to meet these requirements and provable improvements
in performance capability, and regulators want to be able to set thresholds for
performance in different environments that can eventually enable guaranteed

44 of 95
safe implementation of humanoid robots in the highest complexity, but also
highest value, areas (e.g., humanoids walking in complex, crowded spaces
alongside humans, able to collaborate, etc.).

Test Methods and Task-Based Metrics

Standardized test methods present repeatable robot tasks and metrology

procedures for measuring performance. Such test methods detail a procedure
to follow during a test, as well as any physical apparatuses that must be created
to conduct the test. These methods typically focus on a particular task to be
performed, and include any repeatable disturbances or variations in the test
that should be used in the evaluation.

To apply a similar strategy to evaluating the balance/stability of humanoid

robots, we recommend the creation of a suite of stability-related test methods.
The tasks that these methods evaluate should be of significant variety to cover
all the most probable situations that a humanoid robot will face in the
workplace. Variations belong to three categories:

• behaviors to be evaluated, including templates of operations or

stereotypical situations;
• changes in the environment in which these tasks would have to be
performed;
• robustness to common disturbances that the robot might be exposed to
in such configurations.

Table 2 below lists a sample set of test methods that could be developed for a
humanoid robot that would be implemented in simple load pick-up, move, and
place tasks in a standard industrial environment.

45 of 95
TA B L E 2: E XA M PL E T E ST METHODS TO EVA LUAT E
STA B I L I TY I N H U M A N O I D R O B OT S .

TEST METHOD DESCRIPTION METRICS

Robot maintains a static pose

on various surfaces (flat, Maximum disturbance force/
STATIC BALANCE TEST inclined, uneven) under torque, sway angle, time to
external disturbances (pushes, recover balance.
pulls).

Robot walks on flat, inclined, Walking speed, step length/

and uneven terrain, height variability, number of
DYNAMIC WALKING TEST
navigating obstacles and stumbles/falls, energy
maintaining balance. consumption.

Robot carries varying payloads

Maximum payload, stability
LOAD CARRYING STABILITY (weight, size, center of mass)
margin, task completion rate,
TEST while performing static and
fall rate with load.
dynamic tasks.

Robot performs manipulation

Manipulation accuracy, task
tasks (picking, placing,
MANIPULATION STABILITY completion time, stability
assembling) while
TEST margin during interaction,
maintaining balance and
force limits.
resisting interaction forces.

Robot is intentionally
destabilized or falls, and its Recovery time, success rate of
FALL RECOVERY TEST ability to recover to a safe recovery, impact forces during
state (e.g., upright, curled) is fall, final safe pose.
measured.

Robot interacts physically with

Interaction force limits,
a human (e.g., handoff,
HUMAN INTERACTION STA‐ stability margin during
guiding) while maintaining
BILITY TEST interaction, human comfort/
balance and ensuring human
safety ratings.
safety.

46 of 95
When creating these validation methods, the procedures should be flexible
enough to allow robots of different motion capabilities and morphologies. The
tests should be practical and rigorous, easy enough to set up and conduct so
that they can be widely used, but still difficult enough to accurately differentiate
robots of various capabilities.

From each test, numerous evaluation data points can be taken, such as success
rate, speed of execution, accuracy, etc. Each of the individual example tests that
could be created can be considered to be a basic capability. Since a humanoid is
a general-purpose robot, more complex behaviors can also be tested by
combining multiple of the basic capability tests outlined in Table 2. Such
combination tests should be performed if the application requires specific use
cases that can be evaluated.

One limitation of this approach tied to test method performance is that the
total number of possible permutations of tests is very high, more than the
number of test methods that can realistically be written. This is a limitation of all
demonstrated behavior-based measurements of performance. Especially for
humanoid robots, which are performing multiple tasks at once, it has not been
explicitly proven that just because the robot can perform the pre-defined test
methods well, it can also perform well on the tasks in a real work environment.

47 of 95
The correlation between performance on test methods and performance in real
work is still meaningful, and more research should be done to understand the
uncertainty in this relationship.

Beyond template operations to be accomplished, there is also a need to

evaluate intrinsic capabilities of the robot that are not directly measured by the
statistics of application-related scenarios. For example, capabilities like swinging
arms to aid in balance while walking, state estimation of loads when picked up
for manipulation, or replanning footsteps to adjust to disturbances. All of these
properties affect stability, but they are only indirectly captured in basic test
methods. Specific test methods can be created for any special capability of the
robot and should be used to understand how that feature impacts the overall
robot, possibly without compromising any safety constraints.

Control-Based Metrics

While test method performance is the first and most straightforward way of
evaluating humanoid capability, completion of those tasks requires a
combination of control, modelling, planning, and behavior generation
functions. Walking stability is an inherent part of that system, but it can be hard
to isolate the specific stability performance when only evaluating by task
performance. As such, targeted metrics for stability performance need to be
identified or produced that can be implemented during tests. To be
standardized, such metrics should be repeatable, robust, and accurate enough
without requiring too expensive equipment to measure.

Stability and balance control of humanoid robots has been a well-researched

topic for several decades now, and so there are various metrics that have been
previously presented that a standardization process can start with. A few
examples are: the instantaneous capture point (ICP), zero-step-capturability, the
margin of stability, sway angle, or maximum allowable angular momentum.
Most measures of stability, including these, are instantaneous measures of
stability, able to interpret the current state of the robot and identify if it is
currently unstable, or how close to unstable it is. This type of metric is useful for
evaluating the performance of a robot after it has completed an action, as the
instantaneous stability could be evaluated throughout the motion. However, to
meet requirements laid out in current standards, that no motion or action
should destabilize the robot, some amount of prediction of immediate future
performance of the robot, given a planned set of motions, is required, especially

48 of 95
for more dynamic actions such as walking. Many state-of-the-art humanoid
robot control approaches already incorporate some amount of prediction, such
as with optimal control or model predictive control (MPC) based approaches. To
fully evaluate stability performance, metrics will need to incorporate predictions
of possible future dynamic states and what is acceptable as “stable enough”.

The closest current works to forward-looking stability metrics are control

approaches that include “safety” constraints, or dynamic stability constraints in
the motion planner, and in the executing controller. Examples of such
constraints are one-step-capturability and stability regions [2]. These constraints
are based on calculating the regions around the robot during a foot swing
phase in which a robot would be able to stabilize itself if it places its foot down
within the region, and then to always ensure the robot can reach this region.
However, these constraints end up as simply necessary criteria for stability,
rather than a true metric of stability that could be maximized in a motion
planner, or reported as a metric of performance. These stability constraints
could be useful in creating a standard that establishes the minimum criteria
that must be met to be considered a stable action. Such a standard would only
be a first step towards establishing standards for stability quantification and
measurement.

Beyond knowing when a robot is unstable, there is a need to know how stable a
robot is, such that other motions and interactions can be planned accordingly.
For industrial and other applications in human spaces, reliability and safety are
primary concerns for which measures of stability and an understanding of how
stability determines behaviors are needed. Higher stability and more robust
motions may be required, or desired in some situations, while lower stability, but
higher performance actions may be acceptable in others. While a standard
measure of stability is not needed to design the structure or capability of a
robot, it would be extremely beneficial for risk assessment and proper
implementation planning. Additionally, just like humans walking, humanoid
robots in real-world spaces would be subject to disturbances, interruptions, and
other unpredictable events. As such, the goal of these metrics is to have a way
of continually evaluating the performance of the robot, not only during
validation tests, but also during real-world applications as a monitor of
performance.

49 of 95
Finally, an additional need for stability metrics in the creation of future
standards is for realizing safety requirements in non-standard tasks. The
simplest way to prove that a robot can safely perform a task is to establish a
standard test method evaluating the performance of that task, and then
conduct trials validating the required capability. However, it is impossible to test
every action that it is desired for humanoids to accomplish, so there must be
measures of stability that can be applied to any generalized task. Showing
performance with these metrics will then allow for meeting the guarantees of
stability and performance that controllers must deliver, as required by safety
standards for various classes of robots and applications.

Development of Safety Standards

Closely intertwined with the quantification of stability performance is the

development of safety requirements for bipedal robots in various application
domains. Current safety standards are insufficient for fully addressing the
capabilities and operations of bipedal (active-balancing, stepping) robots.
Implementation following current standards would result in a severe limitation
of the scope of potential applications that these robots could have, not from a
lack of capability, but from an inability to determine that the task could be
performed safely. This stems from a lack of safety definitions and requirements,
as well as a lack of methods to validate safety-related control functions of
bipedal-type robots. Such validation can be initially performed simply by
defining performance thresholds on key test methods, which together may
establish confidence in safe operation on known tasks. However, to achieve safe
general-purpose robots, additional safety methods and stability guarantee
requirements will need to be created that can verify the performance of robot
subsystems such as controllers, sensing systems, and internal robot models.
Establishment of clear, well-defined, objective, and easily verifiable safety
requirements is the key for addressing stability concerns arising from dynamic,
legged robots, and thus is a major step towards supporting the successful
development and implementation of humanoid robot technology.

Limitations of Existing Safety Standards

In the current absence of a safety standard for humanoids, current deployment

solutions would need to temporarily reference some elements in existing
standards. This is a common practice in contingent conditions, but it requires

50 of 95
careful consideration of the scope and degree of interpretation of standardized
requirements. In general, safety technical standards must not be used à la carte.
Potential options to consider are the (obvious) complete physical separation
between robot space and human spaces, and the residual exchange of energy
in case of physical contact. In both cases, the impact of the active control of
stability has a direct effect on the size of separation to either maintain from
humans or to evaluate when defining a safe state based on residual contacts.

The most relevant reference is then ISO 13855:2024 “Safety of machinery —

Positioning of safeguards with respect to the approach of the human body”,
which illustrates the methods for computing the safe distances concerning
static and moving sources of hazards, to set or configure protective devices. The
standard is valid across products. In the case of mobile robots, the source of
hazard is moving, together with humans, so all components for safe separation
(relative velocity, time to come to a safe state, distance between potentially
colliding objects) shall be adapted and calculated together. The implications of
maintaining distances propagate to the components that measure distances
during the execution of safety functions and behaviors.

The second reference is the standard for collaborative operations using

industrial robots: ISO/TS 15066:2016 “Robots and robotic devices — Collaborative
robots”. The main clear mismatch for humanoids is that the scope of the
standard is limited to robots fixed on the ground, or mounted on constrained
tracks, or part of an integration onto statically stable bases. If motion is stopped
in actively controlled joints as a result of safety limits, then the robot will not lose
stability, result in falls, or extend beyond the pose at stopping time. ISO/TS
15066:2016 enforces safety limits (a) on the (dynamic) safeguarded space around
a robot to maintain separation or (b) on the contact conditions that may occur
whenever a human is in proximity or shares workspace with a robot. The
modalities are known as Speed and Separation Monitoring (SSM) and Power/
Force Limitation (PFL), respectively. In both cases, limitations on a humanoid’s
motion will have to deal with guaranteeing that the possibility of falling will
reduce the risk to the human to the greatest possible extent. Again, active
control of stability may be high-performing, but not perfect, as robots can still
topple if inputs or disturbances are too large (just as humans can sometimes
fall over).

51 of 95
For SSM-like approaches, mobility and active control of stability to remain
upright will require the introduction of additional space to create a collaborative
non-contact zone. If a human approaches this safe region around the robot,
then the robot could, for instance, enter a stance and control mode that is
practically motionless. For a robot with two legs, that would mean with both
feet on the ground and not taking any steps or at least not any step that could
generate large or unpredictable displacement. Common sense would indicate
that such a robot pose should be such that if power were to be lost (or an e-stop
hit), the robot would fall away from the human, and not towards it. In this basic,
yet robust, implementation of the SSM criteria, the exact definition of the size of
the stable safe region is determined as a design solution (not from the
standard’s requirement). For instance, it could be a region with a radius
matching the height of the robot, and then shaped based on the current
centroidal velocity such that the region represents the possible space a robot
could fall into if it were to topple in any direction at any moment. Note that we
are not necessarily recommending that this is the exact adoption of a standard,
nor should it be the standardized requirements. It is instead an example of a
solution to illustrate what type of framework and requirements (i.e., the targets
for such solutions) are needed in safety standards dedicated to humanoids.

For PFL-like approaches, it’s currently completely unknown what the potential
configurations or limits are for residual contacts within the limits of pain onset
that are specified for collaborative applications. ISO/TS 15066 (resp. ISO 10218-x:
2025) is, in fact, completely dedicated to single manipulators, while the complex
effects of energy (or power) flux density, effect of exposed contact surfaces, and
distribution of forces are unknown for humanoids. Standard methodologies for
recommending the settings of limits and their verifications are heavily restricted
to the hypothesis proper of manipulators. Still, the principle of quantifying the
effects of physical interaction illustrated in the PFL mode of ISO/TS 15066 is
becoming a solid foundation for the extension of such standardized limits to
general machinery when contacts are part of the intended application.
Important to be reminded, the PFL mode involves a purposeful collaboration
between humans and robots in the same shared space. Accidental contacts
have a distinct risk profile that can be studied by the specific workflow of such
applications. An extension to occasional contacts for non-task-related
operations (e.g., random contacts with bystanders) is, in general, out of the
scope. It is indeed tempting to consider collaborative limits generic because a

52 of 95
mobile robot can be anywhere in a working space. However, this shortcoming
would bypass a due risk assessment with obvious consequences in missing
proper estimation and evaluation of risks.

In conclusion, references to existing safety standards that give requirements for

human proximity or potential exposure may be found, but remain largely
insufficient or, at the very least, subject to extensive discussion on the
conditions of applicability. The verification of safety functions dedicated to
implementing the separation or physical interaction principles remain very
challenging because of the lack of clear indication from a functional safety
standpoint. Partial or total removal of humanoids from human spaces will
severely limit the potential capabilities and uses that customers and
manufacturers want to implement using humanoids. Note that this observation
would apply to any robot actively controlling stability, which includes, but is not
limited to, legged robots. There is a need for a safety standard that specifically
details the behavioral and design requirements for a humanoid robot to be
safely implemented.

Safety as Performance Thresholds

To create new standards for safety, especially as it is related to stability, the most
straightforward approach would be to establish performance thresholds on
tasks that the robot needs to be able to perform, and then require robots to
demonstrate that level of capability on those tasks through validation test
methods. The test methods and metrics discussed in section 3 could be used
for this purpose, with the exact threshold that must be met for any particular
metric-task pair determined by the user and application. The set of required
tasks and the level of the safety thresholds can be altered for different
applications and environments. For example, a humanoid designed to go out
into the public may require a higher level of stability performance and
robustness to disturbances, but a lower level of manipulation accuracy,
compared to a robot that is to be used in an industrial environment with only
infrequent interaction with fully trained professionals.

In this approach, a robot would demonstrate the capability to perform a task

safely, and then, as long asmost of the tasks a humanoid is expected to do fall
within the scope of those capabilities, it can be expected that the robot can
safely perform its required tasks under a minimally varying set of conditions. For
example, if safety requires that the humanoid robot be able to fall away from

53 of 95
any nearby humans in the case of a fall, the capability can be shown via a
disturbance-based test method that represents the tasks in question, that a
robot can fall in a predetermined direction if a fall occurs. If, during operation, a
robot enters a state or environment for which it does not have as rigorously
proven stability, then a different, fallback safe state could be utilized. While an
approach like this is useful and would be an accessible first step to establishing
stability and safety requirements, the lack of generalizability of safety to any
particular task is a limitation of this approach. More sophisticated safety
requirements will need to be defined that can handle any arbitrary task a robot
might need to perform.

Going Beyond Current Standards

As stated above, a performance threshold method of proving that a robot can

perform a certain task safely does work well to establish robot capability.
However, it still does not offer guarantees of safety or stability when a robot is
implemented in a real-life environment. A robot may be capable of performing
a task safely, but to meet standard safety requirements, the robot must be
guaranteed to perform the task safely, or it needs to alter its task plan to one
that it can.

As a result, we can state that there are two parts to meeting the safety
requirements to be set out in future standards: first is demonstrating the robot’s
capability to safely perform a task (as measured by specific metrics and test
methods), second is implementing guarantees that the robot will perform the
task in a real-world implementation, subject to any foreseeable disturbances.
Such guarantees could look like stability constraints in the controller, high-
accuracy path tracking, or predictive control that models the robot throughout
an entire dynamic motion. Several such approaches have been presented in
academic research, but to be implemented in practice, new standards are
needed with the details and procedures necessary to validate such controllers.
Additionally, if meeting these safety requirements involves state estimation of
an internal model or external sensing of the environment, then validation of
sensor and modelling reliability will also need to be presented.

These details and procedures are needed to guarantee the stability, and thus
safety, of a bipedal robot within foreseeable and reasonable circumstances.
However, it is recognized that 100% guaranteed stability is not always possible.
Just as humans sometimes will fall over due to unforeseen circumstances or too

54 of 95
large of disturbance, a bipedal robot may fall over as well. This means that
standards for fault handling, fall response, and handling of all other errors must
also be presented.

Finally, a better understanding of the relationship between possible instability

and risk is needed. The risk of a robot falling over scales up many factors,
including the weight of the robot, the mass of the payload, the maximum
energy/momentum that it is allowed to move with, and the environment it is in.
The risk will also change depending on the robot’s robustness to disturbances
and the capability of the balance controller. Standards will need to define
minimum levels of acceptable risk for various application spaces, including
what minimum levels of stability will be necessary to achieve stability with
dynamic interactions. The key point to operate within the same workspace
together with humans is to set and measure stability thresholds that are
deemed safe enough.

Posed in more procedural terms, the tried and true approach to managing risks
inherent to machines in human spaces is to institute appropriate controls that
reduce the risk. However, the behaviors of balance and stability that bipeds
exhibit are produced by increasingly complex algorithms (optimal control,
multi-layer MPC, learned policies) that make it harder to understand what risk
the robot poses from possible instability, and therefore harder to understand
what the appropriate controls are. As the controllers that govern robot behavior
increase in mathematical and algorithmic complexity, the standards that are
used to measure and evaluate these controllers will need to be developed to
match the capability. This does not replace the need and use for control-
independent, task-performance test methods, but should be in addition.

Of note is that requirements listed in future safety and stability standards

should be specific and descriptive, connected to how such robots are designed.
Most current safety requirements for robotics are very loosely defined. For
example, the required capability for performing SLAM on autonomous vehicles
is generally “fit for the safety goal”; the requirement to avoid instability on
mobile robots is open-ended; the ability to identify nearby humans for
collaborative robots is not even close to defining a target success rate. General
requirements are not detailed enough to be useful for humanoid robots and
other complex robotic applications beyond simply requiring “safe” operation
near humans. Definitions of sufficient safety (sufficient levels of risk reduction)
must be agreed upon and set. Additionally, approaches for mitigating risks can

55 of 95
go beyond behavior generation constraints and safety control functions, to
include physical design specifications such as soft contact points or low centers
of gravity and modular safety devices such as air bags or overhead gantries in
hazardous areas.

Overall, future standards will need to do the following:

• reference performance metrics to address the specific capabilities of the

robot in a quantitative form;
• establish targeted quantities that define the level of safety;
• consider foreseeable interactions with the environment and other actors
(both human and robotic) that can modify the threshold quantities;
• define and set the amount of prediction up to a defined time-horizon over
such quantities (thus, standards must be developed to consider internal
aspects of controllers and algorithms);
• define and provide safety validation criteria directly for control, sensing,
and motion planning capabilities, rather than only defining a set of tasks
that a humanoid should be expected to safely perform without losing
stability.

Reaching consensus on the establishment of such standards will necessitate a

highly coordinated effort between manufacturers, regulators, researchers, and
end users. The coordinated efforts of multiple standards development
organizations will be needed to accomplish this goal. As such, the
recommendation is to establish a group dedicated to specific requirements for
humanoids. The overarching standardization goal is to ensure that standards
evolve towards the ideal of objective safety guarantees and test methods, but at
a pace that design and implementation can keep up.

Future of the Study Group

Given that a large standardization effort over an extended period will be

required to achieve needed standards for humanoids, cooperation and
collaboration will be needed between all SDOs, and with manufacturers,
researchers, and customers. To manage such an effort, a central working group
is needed to keep track of the ongoing efforts in multiple SDOs and ensure
research gaps are being addressed. Another recommendation is that this IEEE
Study Group on humanoids continues beyond the publishing of this report to
fulfill this role. However, because a humanoid is a ‘general-purpose’ robot, its

56 of 95
intended functionality extends into all aspects of robotics, as well as the
standards that shape them. Therefore, this group will need to successfully work
with other groups and SDOs, addressing all other aspects of robotics as well.

At the time of the publication of this report, a new project from ISO/AWI for
developing a safety standard that will cover industrial bipedal robots has begun:
ISO/AWI 25785 Part 1: Safety requirements for industrial mobile robots with
actively controlled stability (legged, wheeled, or other forms of locomotion)--
Part 1: Robots. This is the first, and a significant, step towards creating updated
safety standards for bipedal robots.

Conclusion

In this chapter, humanoid robots are characterized by bipedal, powered, self-

balancing operation. This legged nature enables dynamic active control of
stability combined with dynamic reshaping of the robot’s support polygon,
which allows the robot to theoretically navigate any potential terrain while
completing arbitrary tasks. This capability, however, also increases the risk of all
motions due to the potentially unstable status of bipedal legged locomotion.
Humanoids (including both humans and bipedal robots) are always one
mistake or unexpectedly large disturbance away from a hazardous falling or
other hazardous limb motion. The high level of risk means a high validation
threshold to be achieved for stability-related safety control functions. However,
no current standards account for robots that have unstable states as part of
regular functioning. Standards must be extended and/or newly created to
address the unique capabilities of humanoid-type robots. This chapter outlines
a multi-directional approach to the improvement of standards for measuring,
evaluating, and guaranteeing stability, together with standards that lay out the
practical requirements for maintaining the safety of bipedal robots. Specifically
needed are: safety standards that lay out the requirements for maintaining safe
operation in various application domains, standard metrics for the
measurement of stability and performance of bipedal robots given their
complex motions and multi-objective control systems, and standard test
methods to form a common understanding of performance across varying
robots and physical capabilities. Within these standards, measurements of
stability will need to be consolidated in objective metrics, verification criteria,
and validation scenarios. Safety standards will need to provide a dedicated
illustration of risk assessment regarding stability and clear requirements for

57 of 95
safety functions and safety behaviors that would include several contributions
to risk reduction. Safety standards will benefit from testing standards, as most
safety behaviors will require quantifiable conditions to establish minimum
acceptance thresholds of performance. Additionally, as new subsystem
performance requirements are defined (balance, state estimation, navigation,
multi-objective coordination), new standards will need to be created that can
verify the capability and reliability of these subsystems.

Humanoid robots, like all machines, will never be perfectly risk-free. However,
with the proper safety controls implemented, there can still be a safe and
effective implementation of the technology across many intended application
domains. Creating the standards, as described above, will provide the tools and
common understanding necessary to achieve these successes. Finally, the same
standards created with humanoid robots in mind will apply to a wide variety of
robots that share the targeted features with humanoids. Standard efforts
resulting from this report will improve the implementation and applicability of
many robotics technologies throughout industry.

Sources:
https://www.marketsandmarkets.com/Market-Reports/humanoid-robot-
market-99567653.html

https://www.marketsandmarkets.com/PressReleases/usa-humanoid-robot.asp

https://www.snsinsider.com/reports/humanoid-robot-market-1616

https://www.cervicornconsulting.com/humanoid-robot-market

ASTM humanoid robot database

https://interactanalysis.com/insight/humanoid-robots-large-opportunity-but-
limited-uptake-in-the-short-to-mid-term/

58 of 95
Human-Robot Interaction for
Humanoids
Authors: Francisco Andrade Chavez, Benjamin Beiter, Marie Charbonneau,
Brandon J. DeHart, Greta Hilburn, Jeremy Marvel, Kartik Sachdev, and Dieter
Volpert

Humanoid robots operating in human spaces will typically be highly interactive.

In particular, humanoid robots are envisioned as valuable tools for task-oriented
assistance in a variety of applications that will involve interactions with humans
of various backgrounds and abilities, including but not limited to household
chores, caregiving, or healthcare. Therefore, in this chapter, we will focus
specifically on the interaction between humanoid robots and humans.

Although human-robot interaction doesn’t always have clear safety

implications, the collected knowledge and experience that have been distilled
into this chapter will demonstrate that it is crucial to include a consideration of
the interaction aspects of humanoids in any discussion of safety and standards
development, as humanoid robots are inherently “interaction machines”.

Human-robot interaction (HRI) is a broad field of interdisciplinary research

being investigated by a diverse swath of academia, industry, and government
organizations. The primary goal of this research is to better understand how to
develop, measure, and improve interactions between humans and robots as
they become more and more prevalent in our society. Human-robot interaction
begins with the first impression that a robot inspires in a human. Similar to how
humans interact, this initial first impression evolves through further interaction.
These first impressions are primarily based on the appearance of the robot,
including the robot’s physical features, how it moves, how it sounds, its
mannerisms, and how it behaves.

How does a humanoid robot designer know what details to include or change,
and how much? There is much research on these topics, and more needs to be
done as humanoids (as highly interactive robots working in human spaces) are

59 of 95
used more broadly. To start to understand the various aspects of this problem,
however, we can lean on the expertise of those in the field and the expectations
of prospective users, as described in the next section.

Qualitative Data Collection on HRI and Humanoids

To facilitate a deeper look into this topic, a qualitative survey was developed and
shared with stakeholders across many different fields, both within HRI itself and
beyond. The intent of distributing this survey widely was to collect responses
from experts on the topic, experts on related topics, and laypeople to generate
an indicative cross-section of opinions on which to base the discussions and
recommendations found later in this chapter.

At the start of the survey, respondents were informed that its purpose was to
help us determine what aspects of HRI might impact the development and
application of standards for humanoid robots. The survey consisted of the
following 12 questions, divided into three thematic sections:

Robot Appearance and Communication Methods

• How does a humanoid robot's appearance/presence/pose/motion impact

how unfamiliar people react to it?
• How can the appearance/presence/pose/motion of a humanoid robot lead
to incorrect assumptions about its capabilities?
• What should the minimum communication capability/capacity be based
on for humanoid robots (e.g., application)?
• What key elements are required for people to recognize the intent of a
robot's actions or behaviors?

Communication Behaviors

• How should humanoid robots communicate their intentions (upcoming

motions/behaviors) to people in their vicinity?
• How should humanoid robots communicate their awareness of people's
presence/actions/intentions in their vicinity?
• How should humanoid robots communicate potential accidental/
intentional physical contact with the people around them?

60 of 95
 • How should humanoid robots respond to accidental/intentional physical
contact with people to ensure both safety and utility?

Safety and Uncertainty

• What does a 'safe’ or ‘trustworthy’ humanoid robot mean to you?

• What should the minimum safety-specific behaviours be of a humanoid
robot depending on its application and environment?
• How should a robot balance risk to itself and other equipment with risk to
nearby people when planning dynamic motions and tasks in its
environment?
• What should the role of a human supervisor be in applications and
environments with different amounts of uncertainty?

Over a few months, we collected fifty responses to this survey, with responses to
each question from each respondent ranging from a few words to several
paragraphs. Based on the inductive method of qualitative content analysis
process introduced by Elo and Kyngas in 2008, we have summarized and
distilled the responses to inform the discussions and recommendations below.

User Centered Design Supportive Research

This chapter has also been informed by a set of 12 user interview participant
responses collected regarding the potential integration of robots into daily life,
as part of a study conducted in the USA by one of our team members. This
consisted of an in-depth analysis of 12 user interviews on humanoid robot
integration that included a diverse range of participants (all non-engineers),
which reveals a complexity of expectations and ethical considerations that are
critical for future development.

The user interview questions focused on user expectations as well as current

general perceptions concerning humanoid robots, to support the HRI survey
and find any potential missed opportunities in design, safety, and future user
expectations. Participants articulated viewpoints ranging from profound
concerns about the ethical implications of robot "slavery" to optimistic visions of
robots as compassionate caregivers and assistants.

61 of 95
The findings consistently underscore the importance of emotional intelligence
and trust in human-robot interactions, with users emphasizing robots' need to
comprehend and respond to human emotions, especially when engaging with
children, older adults, and other vulnerable populations. The user interview
responses regarding accessibility, data privacy, and potential social implications
highlighted an urgent demand for ethical design and equitable deployment.

Regarding humanoids in human spaces, there's a clear desire for robots to be

task-oriented and functionally reliable, but users also expect personalized
interaction with the robots, including meaningful communication. A prominent
concern emerged around robots becoming a luxury status symbol, advocating
for business models that prioritize affordability and equitable access through
mechanisms like rental programs or insurance.

Ultimately, this user feedback points to the necessity of a human-centered

approach to robot development, one that prioritizes safety, simplicity, and the
cultivation of positive human-robot interactions based on positive connections.
The core findings emphasize that for humanoid robots to be safely and
effectively integrated into daily life, particularly for vulnerable populations like
children and older adults, they must prioritize emotional intelligence,
trustworthiness, and seamless accessibility. Users expressed a strong desire for
robots that demonstrate empathy, understand human emotions, and build
genuine rapport.

This user interview feedback informs standards by highlighting the need for
humanoid robots to not only avoid physical harm but also prevent emotional
distress, ensure data privacy, and mitigate potential social disruptions. It also
outlines that the end user will be expecting: (i) a focus on accessibility and
inclusivity, (ii) emotional intelligence and understanding to be prioritized, and
(iii) seamless human-robot communication.

Similarly to the survey feedback, users wanted a sense of genuine

understanding between themselves and the robot, although the user interview
feedback included a need for more empathetic interactions, including:

• Human-like communication: beyond simple speech, the robot should

mimic natural conversation patterns, including pauses, response
anticipation, and empathetic tone.
• Non-verbal communication: eye contact and realistic hand gestures are
crucial for conveying warmth and attentiveness.

62 of 95
 • Accessibility and inclusivity: the robot must be capable of
understanding and responding to users with communication
impairments, such as those caused by strokes or other health conditions,
as well asl children who have suffered trauma with trust issues.
• Emotional intelligence: the robot should be able to interpret subtle cues,
including minimal facial expressions (micro-expressions) and vocal tones,
to understand the user's emotional state and needs. Examples given: fear,
pain, confusion, and trust.

Supporting the HRI Framework

The user-centered design recommendations based on the user interviews,

particularly those on advanced emotional connections, are foundational to
establishing trust and perceived safety. A robot that can interpret and respond
empathetically is inherently perceived as safer and more reliable, reducing
anxiety and increasing acceptance.

Furthermore, the proposed real-time trust measurement and renewable

licensing model that were discussed in the user interviews offer innovative
approaches to continuously monitor robot performance against established
safety and ethical benchmarks. This dynamic feedback loop can inform
regulatory adjustments, incentivize manufacturers to prioritize human well-
being, and provide a concrete mechanism for ensuring robots meet evolving
safety and social standards.

In essence, this qualitative data argues that effective safety and standards for
humanoid robots cannot be developed in a vacuum. They must be deeply
informed by real human needs, fears, and aspirations, ensuring that technology
serves humanity in an inclusive, empathetic, and ultimately, safe manner.

The remainder of this chapter consists of 3 main sections, each dedicated to one
of the thematic topics included in our survey within the field of HRI as it relates
to humanoid robots, followed by a section outlining the lessons learned via the
user interviews mentioned above, and concluding with a section outlining our
overall recommendations for standard makers and robot designers, based on
both internal expertise and the results of the survey and user interviews.

63 of 95
 Robot Appearance and Communication Methods

To optimize the safety and efficacy of human-robot interactions, designers must

consciously prioritize a robot’s appearance (along with its dynamic motions and
physical capabilities), as it dictates how users will likely perceive and interact
with a given robot. On this topic, the survey asked how respondents think a
robot’s appearance can impact communication between a user and the robot:
how it shapes first impressions, the assumptions it can lead to, and what
specific features and capabilities they think humanoids should, and will, have in
the future.

Broadly, the responses we collected confirm that appearance is indeed

important for shaping how humans and humanoid robots communicate. Users
set their initial expectations of robot capability based on the appearance of the
robot, and especially on how similar it looks to a human. This appearance-based
first impression can also determine the levels of comfort and familiarity a user
has with the robot, even after interacting with the robot regularly.

While both factors can be used to convey necessary information quickly and
intuitively, mismatches between the impressions given by appearance and the
real capabilities of the robot can cause safety concerns. People who feel too
familiar with the robot, or think it has a higher cognitive capacity than it does,
will be more likely to not show proper caution and to not follow proper
procedures to keep themselves and their surroundings safe. Such concerns can
be mitigated by proper training, signage, and other controls, but the
appearance, actions, and direct communication of the robot with the people
around it will continually reinforce certain impressions, so designers should be
deliberate about conveying accurate information.

Based on user responses, three types of communication may be needed and/or

desired from the robot: Explicit, Implicit, and External, as described in
Communication types. While users have varying opinions about what the
minimum threshold of technical communication capability or modality should
be, Communication Purpose and Means describes how the purpose of
communication with a humanoid robot was considered to have three
components. The first is to convey the robot’s intent, allowing nearby humans to
accurately predict the robot’s next actions, while the second part should convey

64 of 95
the general state of the robot. Finally, the third component conveys details
about the environment a robot is working in, the obstacles it needs to work
around, and tasks that the robot is performing.

Building on this 3-part perspective on human-robot communication, we will

next explore how the relationships between robot appearance, communication
types, capabilities, and human expectations need to be accounted for in the
design of a humanoid robot.

Appearance and Capability Alignment (Emotional

Expectations)

Robot appearance significantly influences user expectations and trust. Accurate

alignment between appearance and actual capabilities is crucial to avoid
misinterpretations that could lead to safety hazards or diminished trust.

With one exception, all survey respondents agreed that appearance does have
an impact on communication. While this is potentially helpful, it can also lead to
miscommunications when humans make false assumptions about the
capabilities of a robot based on its appearance. Many noted that the more
human-like a robot looks, the more people will expect it to be able to perform
actions (both in mobility and communication) equal to what a human can do.
Based on user feedback from the user interviews, the majority of individuals
who were interviewed agreed that a robot's appearance should be determined
by the robot's intended function.

As some survey responses noted, anthropomorphism is inherent to being

human, as we apply it to everything, but humanoid robots convey
anthropomorphism on another level, even prompting character design, giving
the robot an identity to interact with rather than just base functionality.
Conversely to anthropomorphism, if a robot has more ‘robot-like’ features (flat
surfaces, hard shapes, points, jerky motions, etc.), then people will tend to treat
it more like an autonomous robot, expecting less interactability or intelligence
and more machine-like behaviors.

When evaluating the effect of appearance, the most common response was
that, currently, the anthropomorphic appearance of humanoids across the
board causes people to overestimate the robot’s true capabilities. People
assume it has mobility beyond what is possible at the moment, a level of

65 of 95
sensing and awareness beyond reality, higher intelligence, reasoning, and
emotional understanding than it does, higher reliability in task performance
than currently has been proven, and a higher level of safety in the robot’s
immediate vicinity than is guaranteed. A less common, but still possible
problem is that a less human-looking robot could lead to users underestimating
what the robot is capable of, though this poses less of a safety concern.

Further complicating the matter, everyone comes to an interaction with

different expectations, impacted by prior experience, knowledge, identity,
physical size (if a robot is larger/smaller or heavier/lighter than they are), and
context. These prior expectations then combine in complex ways with the
appearance-based expectations in a variety of ways, meaning that it is
impossible to perfectly define public expectations of robots (which are not the
same as the expectations of trained personnel), though appearance can
attempt to guide it in certain directions. As a result, the physical features and
mannerisms of a robot need to be carefully considered based on the expected
users, and should not just be an afterthought of design wrapped around a
machine.

User feedback indicates a desire for humanoid robots with the ability to be
personalized by user preference and adapted to a diverse range of users by age
and capabilities. The majority of the groups that expressed a desire for a more
detailed appearance in specific aspects of the robot were young children and
older adults. Facial features that could be expressive and could emote directly
were favored for both groups. Options for both digital and more humanlike
features were considered best for the desired features in terms of interaction
based on individual needs.

In terms of capability, interviewed individuals also expressed that the

appearance of the robot when in a caregiver capacity should have more
humanlike qualities, with its arms and hands showing empathy in its
movements. Social robots were also indicated as requiring more humanlike
features or having the option to be upgraded or expanded in their capabilities
and appearance.

Many of the interview respondents gave emotionally driven responses based on

real-life situations related to the socioemotional components of empathetic
interactions in specific environments. Home, school, and healthcare facilities
were all specified as places respondents would expect to see a humanoid robot

66 of 95
that had features that would signal their abilities to users. These specific
features would be indicators included in its design deliberately to visually
display its purpose, functionality, and capabilities to all humans who see it.

Expectations of the robots' mobility and physical gestures were also emotionally
driven. If the robot was intended for basic domestic assistance, individuals
wanted to have the ability to personalize the robot’s appearance to match the
environment, and have the option to “put it away”. Other interview respondents
wanted a high-performing surrogate human companion robot that served as a
caregiver with the ability to interact and move with human precision.

Trust measurement based on how interview respondents explained a

connection to a robot’s appearance was based on Attractors and Detractors.
Attractors were in the comfort that respondents felt from the visible
understanding of a simple robot design. Detractors were a respondent's
emotional conflict or confusion created by a complicated design in the robot's
appearance. All related to how one might communicate or interact with the
robot.

This also implies that as a robot is continually developed and its capabilities are
improved, the appearance should also change, although implementing this is
not a simple task. Design and movement remain critical in managing
expectations throughout the development of a robot.

Appearance Determines Comfort and Trust

As discussed above, the appearance and behaviors of a robot are a form of

technical communication, implying details of the robot’s capabilities and
competencies. The same features of the robot also evoke an emotional
response from the human, shaping feelings of familiarity, comfort, and trust
within the user towards the robot.

Any feature of the robot can have these effects, including size, color, shape,
voice, mannerisms, motion style, similarity to a human, etc. Many survey
responses associated certain types of features with whether they would evoke
positive or negative reactions in users, and what they believe those reactions
might be, as summarized in Table X. While these features and reactions seem to
be common, the magnitude of the reactions, as well as which features will lead

67 of 95
to which reactions varies greatly from person to person. As with prior
expectations, how an individual reacts is heavily dependent on background,
culture, experience, and many other factors.

TA B L E X: S U M M A RY O F A F F E CT IVE R E ACT I O N S TO T H E
R O B OT A N D WH I C H F E ATU R E S CO U L D B E T H E CAU S E .

POSITIVE NEGATIVE

Calmness, smooth and Aggressiveness, jerky or

predictable motions, cuteness, unnatural motions, overly fast
FEATURES THAT EVOKE THIS
familiar features such as eyes, motions, intimidating form
REACTION:
friendliness, open/natural (size, shape, and hardness), or
poses actions

Fear, disinterest, discomfort,

Familiarity, willingness to avoidance, negative
THIS REACTION RESULTS IN:
interact, and being accepting associations, uncertainty/
wariness, surprise

Communication Purpose and Means

While the appearance and other features of the robot may primarily shape the
first impression a user has of the robot, other forms of voluntary communication
also shape the human-robot interaction. Based on the survey responses, there
were three general reasons identified for why communication between the
robot and human is necessary. Communicating the robot’s intent to the human
is the most commonly stated purpose for communication. A secondary purpose
of communication is to always make a human aware of the state of the robot.
This includes the task it is performing, as well as its operating status, power
status, controller state, health, environmental concerns, and other errors or
faults. Similarly, a third purpose is to communicate the details of the
environment and task that the robot is operating in, including awareness of the
presence and actions of humans in its surroundings.

The form of communication can vary, with communication behaviours that

humans may expect from humanoid robots including visual cues (lights, gaze
direction, facial expressions, gestures or display screens), audio cues (verbal
speech or electronic sounds such as buzzes, beeps and boops), or physical cues
(touch, modulating movements). Overall, given their sophisticated nature,

68 of 95
people may expect humanoid robots to display communication behaviours that
are at the intersection of machine and human capabilities. Further discussion
on this theme can be found in the Communication Behaviours section.

The strategy to generate the communication also varies, from the ability to hold
a basic conversation, to just confirmation of commands when received or
interjection when safety is a concern, such as shouting “STOP”. The overall
purpose is for nearby humans to have a clear understanding of what the robot
is about to do. Essentially, every situation has an associated type and level of
communication that is required to safely perform the task, and the robot must
meet this requirement. Additionally, the communication capability changes
further based on whether the task is in a domestic, medical, industrial, service,
or public environment. Small details such as color, surface finish, lights, etc., can
indicate the type of application that the robot is probably going to be used for
(and, ideally, has been designed specifically to fulfill).

Communication Types

To achieve the goals of the three purposes of communication stated above,

responses detailed many ways of communication projected to be useful for
human-robot interaction. These have been categorized into three main types of
communication: Explicit communication through words or motions, implicit
communication through the shaping of all behaviors, and external
communication through user interfaces (UIs) not located on or run by the robot.
The details of these types are given in the table below.

69 of 95
TA B L E X2: S U M M A RY OF H U M A N - R O B OT
CO M M U N I CAT I O N TYPE S

COMMUNICATION TYPE DESCRIPTION

Capabilities or actions that are done

exclusively for communication. This includes
lights or screens, voice and other sounds,
EXPLICIT:
facial expressions, or simple motions such as
pointing with hands, looking in a direction
with the head, or the pose of the body.

This involves shaping other behaviors to

convey a certain impression. This includes
movement profiles, patterns of motion, gaze
IMPLICIT:
and eye position, body language, posture, and
other nonverbal behaviors that are ancillary to
the functional purpose of the motion.

These are forms of communication that are

not on board the robot itself. This includes
user interfaces, external displays of sensor
EXTERNAL:
information, network communications,
guides, and other training materials to
understand the robot.

Minimum Communication and Emotional Intelligence

Capability Requirement

Within these types of communication, survey responses also detailed the

desired minimum levels of competency believed to be necessary for
implementing humanoids. The most commonly stated requirement was that
the robot look similar to a human and have a human-like communication
capability. Specifically noted was the capability to understand human speech,
gestures, and mannerisms, be able to communicate back in the same way by
holding a basic conversation, and be able to remember individuals to
personalize interactions. Varying levels of technical competency were also
mentioned, from “Better than Alexa” and “The communication capability of a
dog” up to “able to utilize a small task-based vocabulary of interpretable
commands”, “able to understand natural language via NLP or ChatGPT”, and
“able to interact with a central warehouse management system”.

70 of 95
While most of the survey responses specify how the robot should communicate
with humans, several people also state that the robot should be able to interpret
commands from humans as well and be able to respond to them. Robots
should also have an accessible way for nearby humans to halt or stop
undesirable behavior. Based on these responses, users generally expect
humanoid robots to exhibit emotional intelligence, including recognizing and
responding to human emotions, and to engage in meaningful, human-like
conversations. Therefore, advanced multimodal communication (visual, audio,
physical) is essential for conveying robot intent and status, and for facilitating
natural interaction with surrounding humans in the environment.

This is confirmed by user interview feedback, with an expectation of a high level

of emotional intelligence and advanced abilities to communicate and converse
with a robot via voice activation and an expectation of intuitive language
adoption. This was introduced by users both in the preprogramming of the
robot and in the trainable language and task orientation stages of robot
learning, based on individual preferences and needs.

A portion of the communication would likely be performed via a dashboard that

can take in data and share it with individuals and others based on approval by
the user. A therapist, teacher, doctor, or other healthcare provider could have
access to both the data and the potential modifications in behavior required to
make adjustments to the operation of the robotic system.

In interview responses, the desired communication via interaction was often

based on the type of robot in question. If it were a social robot, it would be
dependent on the individual, age, and cognitive abilities, designed to be used in
tandem with a support system of humans. The communication was primarily
verbal and included digital facial expressions and sounds to communicate to
the user, as well as voice and tablet or an alternate device to communicate
feedback.

Robot emotional understanding was also an expected language adaptation

between the user and the robot, especially with extended or repeated usage. As
an example, cognitive barriers were brought up by respondents as an example
of how a robot would have to be trained to provide personalized interactions
with vulnerable populations like children or older adults.

71 of 95
 Discussion and Recommendations

Overall, there seems to be a consensus that the appearance of a humanoid

robot has a large impact on shaping its interaction with humans, which starts at
the first impression and continues to grow and change throughout all
communication. The appearance of a robot also shapes the expectations of
what that robot is capable of.

When manufacturing a robot, it is critically important that the capabilities

implied by a robot’s appearance match its actual capabilities. Mismatches
dramatically increase the risk of harm during interactions. However, due to the
inter-individual variability of reactions to a robot’s appearance, designing a
proper appearance is a complex, difficult problem. A feature that makes one
person trust a robot might make another mistrust the robot, and both could be
wrong, possibly unsafe expectations relative to what the robot’s actual
capabilities are.

Additionally, over- or under-estimation of capabilities is based on the current

public perception of the capabilities of humanoids, and the relationship
between expectations and reality will change as robots’ capabilities grow.
Therefore, manufacturers should carefully consider the appearance of any robot
they design to accurately represent the current capabilities of that specific
robot, with a need to balance features that make others more comfortable and
not scared of the robot, while remaining cautious enough to maintain safety
procedures.

Building on this, it is worth noting that many responses set extremely high
expectations of the minimum levels of required communication capability.
People expect humanoid robots to be able to communicate at the same level of
competency as humans, including being able to understand natural language
and gestures, and then be able to reproduce both by themselves.

Even more so, some responses specified that robots should be able to know
everything about what actions are being performed, be able to reason at a high
level about why it is doing it with “sound logic”, and then be able to
communicate that reasoning if asked. While it is true these capabilities would
be ideal for clear communication, they may be beyond where the minimum

72 of 95
requirements should be set. Even the same responses that detail how people
have too high expectations for humanoid robots still set high expectations for
standards for the robots.

Though responses agree that appearance heavily influences levels of comfort

and trust, people disagree on whether certain appearances have a positive or
negative impact. Some say that human-like appearances increase comfort and
intuitive interaction, while others say that human-like appearance heightens
discomfort because of the uncanny valley and should be avoided. This question
of whether or not a robot should look more or less like a human has no
consensus, so at minimum, manufacturers should consider their prospective
users and which effect is a priority, or the lowest risk, for their application.

All communication and appearance functionality should be defined and/or

guided by consensus standards, with validation by extensive test methods. To
achieve these standards, however, there needs to be an industry-wide push to
identify what features cause what reactions, and how to create humanoid
morphology and the capability to optimally interact with human users. This
includes defining minimum communication capabilities, methods, standard
signalling and recording, safety requirements, and minimum performance
thresholds.

In summary, appearance should be shaped to accurately convey the capability

of the robot, balancing the need to evoke comfort and trust in humans nearby
(or at least avoiding discomfort) while still encouraging proper safety
procedures. Similarly, for communication, the quantity, quality, and mode of
communication should be tuned to efficiently and effectively convey the
necessary information without overcomplicating the resulting interactions.

Communication Behaviours in Social and Physical

Human-Robot Interaction

Communication with nearby humans may not necessarily be critical in settings

where a robot follows a predetermined work sequence. However, projected uses
of humanoid robots will certainly fall beyond strictly predictable contexts. As
introduced in Communication Purpose, humanoid robots are expected to
communicate with human users in different ways to inform users of the robot’s

73 of 95
intent, state, task, and perception of the environment. This indicates that the
majority of users want to be aware of, or able to provide consent for, upcoming
robot actions.

When robots and humans share physical space, communication abilities may
be especially critical for humanoid robots, for example, when communicating
robot intentions (such as upcoming motions and behaviours) and a robot’s
awareness of people’s presence in its vicinity, as well as its understanding of
their actions or intentions.

It is generally understood that accidental contact between robots and humans

should be avoided as much as possible. For humanoid robots operating in
settings where maintaining physical separation from humans is not always
feasible, there will be a possibility of contact between people and robots. In
these cases, a robot may be expected to communicate potential accidental or
intentional contact with people, as well as to respond to contact.

In settings such as healthcare, service, or home environments, humanoid robots

will often need to operate in proximity to untrained individuals. In these cases,
some form of generally understandable communication of intent would be
advisable. It would need to be as clear as possible, so that the robot does not
cause unintended confusion, surprise, or stress in people, but instead that its
movements and behaviors can be easily anticipated by people.

The level of communication required will depend on context and risk

assessments, such as an understanding of who the robot is working in the
vicinity of, what they are doing, where their attention is, and what they can
effectively perceive. Additionally, the communication behaviors and emotional
intelligence expected from a humanoid robot will likely differ depending on the
application and level of proximity between a human and a robot, such as in
situations where physical distance can be maintained versus those where
physical contact may occur.

In all cases, however, to convey meaning intuitively, humanoid robot designers

should aim for communication that is as human-like as possible.
Communication with untrained individuals could also be facilitated by a robot’s
physical appearance if it is designed to indicate the robot’s intended task.
However, the use of multiple modalities (ideally, as many as possible) is critical
to ensure inclusive communication with individuals with diverse sensory and
cognitive (dis)abilities.

74 of 95
Given how human-to-human communication can sometimes be confusing,
how a humanoid robot is designed to communicate robot intent, awareness of
humans, potential contact, and to respond to contacts (as focused on in the
survey) needs to be carefully considered to ensure clear communication. To that
point, survey respondents suggested a simultaneous combination of visual,
audio, and physical communication modalities. In effect, using multiple types of
signals in coordination may help produce clear communication.

Survey respondents also indicated a strong expectation that the way humanoid
robots are made to communicate with humans would borrow from typical
human-to-human communication behaviors. However, they did not exclude
the use of machine-like communication behaviors. For example, colored lights
or beeps from a robot may not always clearly convey the intended information
to untrained people on their own, but they could be part of a robust and
successful communication system when combined with other behaviors.

The following sections break down survey respondents’ feedback related to

each of the four communication behaviour scenarios identified in the survey
questions, which included an increasing degree of physical proximity and
contact with people in the environment.

Communicating Robot Intent

When it comes to communicating robot intentions (including upcoming

motions and behaviours), respondents suggested the use of various visual,
audio, and physical cues.

Visually, lightsignalling may be useful (perhaps in a similar way as car lights

indicate the intended direction of motion, or in a similar way to position
lighting, projecting lights where the robot is intending to move), but standards
may need to establish clear meanings for light signals. Borrowing from human-
to-human communication behaviours, predictable robot motions, as well as
human-like gestures, gaze direction, and facial expressions, may help to
intuitively communicate intended robot actions. Examples may include
pointing, leaning, or turning the head of the robot in the direction of intended
motion. In a more machine-like approach, visual display screens may be used to
display arrows or descriptive text indicating current and intended robot actions.
However, visual cues assume that humans in the vicinity of the robot can see it,
and that the robot has their visual attention.

75 of 95
Aurally, speech (potentially in conjunction with motions) may be used to provide
awareness of upcoming robot actions to nearby humans (e.g., “on your left”), but
it could also be used to seek consent or to negotiate actions with humans, given
the context (e.g., “May I move ahead?”). However, communication through
speech may not always function in loud environments or when communicating
with humans who have hearing impairments; additional communication
modalities would be beneficial in these cases. Survey respondents also
suggested that beeps or tones could be used before each movement to provide
awareness of upcoming motions, provided they do not negatively impact or
cause fatigue in the people the robots interact with.

Physically, slowing down robot motion around humans is often done to ensure
safety, but may come with a tradeoff on productivity. As suggested by some
respondents, perhaps beyond communicating intent, robots could also
communicate to surrounding humans how to move around them, to minimize
slowdowns. Haptic wearables, for example, may be used to enhance human
awareness of robot actions (e.g., vibrating when a robot is nearby).

Communicating Robot Awareness of Nearby Humans

Communicating a robot’s awareness of people’s presence, actions, or intentions

in the robot’s vicinity requires more complex communication than just
communicating robot intent: people may want to confirm whether a robot is
aware of them when nearby. Intuitive, human-like behaviour is likely to be
expected: robot designers should take inspiration from how humans negotiate
personal space and turn-taking to inform their designs’ behavior. Survey
respondents suggested that awareness may be communicated in the form of a
greeting or short message (verbal or nonverbal). However, they also indicated
that the amount of information communicated should be appropriate, as over-
communication may cause confusion or discomfort.

Lights can be used as a visual indicator, but just as for robot intent, standards
may need to be defined. Lights may change color when a person is detected in
the vicinity; they may be used to indicate that the robot is listening and
processing, or they may be used in other ways. Body language, including facial
expressions, gaze direction, and gestures, may be used to indicate awareness of
the presence of humans in the form of a greeting, for example, with a smile, a
brief look toward a person, a wave, or a head nod. Communication may be more

76 of 95
involved, for example, directing the robot's head and gaze towards people, to
show awareness of their presence, and then tracking their movements to
communicate awareness of people’s actions. A display screen could also be
used to inform people of the robot’s awareness. Robot behaviours may,
however, be adjusted to avoid generating feelings of unease in humans who
may feel observed if a robot is “staring” at them.

Speech in natural language could be used to provide verbal acknowledgement

or confirmation of awareness to humans (e.g., saying “Hello” or “Hi” as a
greeting, or “I see you”), as well as to confirm a robot’s perception of a person’s
actions or intentions (e.g., “Are you picking up that box in the corner?”).
However, respondents indicated that just as for robot gaze, the use of voice
could potentially cause unease, and thus might need to be socially adjusted.
None of the respondents indicated that beeps or tones would convey the
required information in this case, but this could be investigated as a potential
new type of standardized robot “language”.

Robot motions can also be adjusted to indicate awareness of humans in the

vicinity, for example by pausing or stopping when someone enters the robot’s
“personal space”, moving more gently or safely around humans, and reactively
adjusting proximity to people.

Communicating Imminent Contact

A humanoid robot’s communication of imminent contact with a human and

response to a contact that has occurred is critical to both safety and utility.
However, before physical contact may be permissible, robots must first be
verified to have the capacity to: detect the proximity of a human, sense contacts,
apply physical avoidance approaches, and physically respond to contacts (e.g.,
by following collaborative robot standards). However, by their nature and typical
use cases, humanoid robots may be expected to approach and respond to
contacts differently than collaborative robots, especially when being used in a
wider range of environments.

In such scenarios, communication should be clear, loud, and obvious to ensure

humans are aware of the situation. The more dangerous a robot is, the louder
and more attention-grabbing this communication should be. Multimodal
communication should be emphasized to ensure that humans get the
message, regardless of their sensory (dis)abilities. In cases where humanoid

77 of 95
robots are interacting with untrained individuals (or trained individuals who
may not be relied upon to remember the training received), human-like robot
behaviours may be more conducive to providing effective communication. This
would help ensure that communication signals are designed to clearly and
intuitively convey an intended message (as opposed, for example, to having
several warning lights and buzzers going off without an obvious meaning).

When interacting with humanoid robots, people may unconsciously expect a

minimum level of human-like deference from these machines. Local cultural
norms and specific contexts may thus need to be considered. For example, it
may be acceptable in certain cultures for a robot to make contact with a human
to prevent potential injury or to navigate a crowded environment, but perhaps
not for other reasons. In some cultures, it may be customary to apologize or ask
for consent before making intentional contact, or to apologize after making
accidental contact with someone. These cultural norms must be integrated into
general-purpose humanoid robots.

While it is becoming well established that humanoids need reliable

communication abilities, physical human-robot interaction is still in its early
research stage. Additional research is needed to appropriately understand
physical and psychological safety needs in these situations. For this reason,
recommendations can be expected to evolve as new findings emerge.

If a humanoid robot is programmed with the ability to detect that accidental or

intentional contact with a human in its proximity is about to occur, it may also
be programmed to alert and provide adequate warning to the concerned
human(s), as a way to mitigate physical and psychological safety risks arising
from the interaction. A combination of visual and audio signals may be used to
get a human’s attention while ensuring that the trajectory of the robot is
predictable.

Visual signals may include the use of light indicators, perhaps following
industry-standard safety color codes for ease of interpretation. Intensity, color,
and flashing frequency of lights may be modulated to convey urgency.
Additionally, robot gestures may be used, such as deliberate movements that
convey the need for caution (e.g., raising the hands, turning the head to point
the gaze towards a potential contact location) or that convey the intended robot
motions and pathway (as discussed in Communicating Robot Intent). Robot

78 of 95
facial expressions could help convey the need for caution (e.g., moving
eyebrows, eyes, and mouth to display surprise). A screen display may also be
used to communicate an intended contact via text.

Audio signals may include the use of speech or electronic sounds (such as
beeps, boops, and buzzes) to warn humans of an impending contact or to
communicate an intended contact. In both cases, volume, pitch, and speed may
be modulated to convey urgency.

Communicating After a Contact has Occurred

An appropriate robot response to contact is critical to ensure safety and utility,
as well as to maintain trust in humanoid robots. Humans may expect a
humanoid robot to respond to accidental or intentional contact in different
ways, including (i) acknowledgingthe physical contact, (ii) assessing the
situation, before (iii) responding. The robot may respond either by stopping
motion, disengaging, or moving back, mitigating the effects of the intended
contact, safely continuing a previous action, or asking for human assistance,
reporting accidents, or calling for backup (be it a person in charge or
emergency services).
In this perspective, survey respondents expressed that communication after
physical contact may be used to acknowledge the contact, apologize for the
contact, gain context from a human, or provide context to a human. For
example, communication to gain context may be used to confirm: (i) the
safety of those involved, (ii) whether there was perceived or actual harm
resulting from the contact, (iii) the intention of someonemaking contact with
the robot, or (iv) confirm the next appropriate action for the robot.
Communication to provide context may for example be used to: (i) confirm
whether the contact was intentional or not, (ii) describe the intent of the
robot, (iii) describe the actions of the robot in response to the contact, or (iv)
communicate the robot’s intent to correct, adapt, or avoid future contacts. To
communicate such information, a combination of visual, audio, and physical
signals may be used.

Visual signals may include the use of lights, although they may have a limited
ability to convey a required meaning when used on their own, as discussed in
previous subsections. Robot gestures and facial expressions may be used, such
as deliberate movements that convey acknowledgment of or apologies for the

79 of 95
contact (e.g., turning the head towards the location of the contact, moving the
hands apologetically). A display screen could also be envisioned to provide
acknowledgments, apologies, or context to a human through text.

Audio signals may include speech, such that a contact may be acknowledged or
apologized for, and such that context may be gained or provided through verbal
communication. Electronic sounds may also be used, although they may have
limited ability to convey the required communication when used on their own.

Physical signals may include communicative robot motion, such as modulating

motion direction, velocity, or compliance (e.g., stopping, slowing down, moving
away, increasing compliance while in contact) to help ensure safety and
haptically convey acknowledgment of the contact. Some survey respondents
also indicated that a physical interface, with physical interaction capabilities
(e.g., a button, a computer, or making use of haptic robot capabilities), could be
used to gain context, for example, allowing a human to adjust robot motions.

Discussion

Among all scenarios discussed above, a large proportion of respondents

suggested having the robot communicate through speech. This could indicate
that humans intuitively consider speech as an effective means to convey
information, potentially due to its ubiquitous use in human interactions. Lights,
body language, display screens, and electronic sounds may be considered as
complementary tools to enhance the clarity and effectiveness of robot
communication. The selection of communication modes should be context-
dependent, for instance, considering who may be interacting with the robot,
environmental conditions, and tasks being accomplished. Additional research
may be needed to improve understanding of personal space, comfort, trust, and
perceived safety in social and physical human-robot interaction, but existing
literature on user experience design and human-robot interaction may provide
a starting point.

Nonetheless, it remains clear that expectations of the interactive abilities of

humanoid robots are highly sophisticated. Survey responses indicate
underlying assumptions that humanoid robots have sensing, perception,
reasoning, and communication abilities similar to those of a human. User
interviews add to that, expressing a need for intuitive and simple user interfaces

80 of 95
that facilitate clear communication and that can be adapted to different users
(offline and online), while including options for plain language explanations
through speech and/or digital display.

In addition, communication mores may differ from one application to another,

for example, with specific jargon, sounds, and gestures used to communicate
between individuals of a given workplace. This will need to be considered in the
design of interactive humanoid robot behaviors. As communication heavily
depends on human detection, it will also be critical to define how the reliability
of mechanisms for human detection can be validated, and how potential
failures may be handled. Additionally, as clear and effective communication
may not always be guaranteed to prevent incidents during humanoid robot
operation, the next section covers safety considerations for human-robot
interaction.

Safety & Uncertainty in Humanoid Robot Interaction

Any given robotic system intended to be used by or around people must exhibit
safe behaviors while in operation, and humanoid robots may come with their
own set of challenges due to their interactive nature, along with their high
complexity, mobility, and power.

The survey invited input on factors impacting the safety (or the perception
thereof) of robots performing tasks with or around people. The survey did not
provide prompts that encouraged respondents to consider safety in any
particular way, but instead encouraged feedback based on their expectations of
safe interactions. Responses were largely focused on physical safety and
touched upon topics including the definitions of safe behaviors, the
characteristics of robots that promote trust, and the reduction of risk due to the
design, functionality, and operating environment of robots. That is not to say
that psychological safety in these interactions should not be considered. Rather,
it indicates a blind spot in how most people think about safety around robots,
and calls for deliberate attention to be brought to this aspect of safety, including
in the topics covered below: how to define a safe and trustworthy humanoid
robot, what behaviours make a humanoid safe for HRI, and the implications of
human supervision.

81 of 95
 Defining a “Safe” Humanoid in the HRI Context

Regardless of whether a humanoid robot is public-facing or merely working in

an environment that is (or was previously) intended for human presence, the
robot must adhere to certain safety guidelines and design considerations. As
emphasized by survey respondents, a safe humanoid robot must be designed
to minimize the risk of physical harm to humans, whether nearby humans
behave predictably or not, either directly or indirectly through its actions and
presence in the environment. This is achieved through a combination of
elements.

For humans to be safe in the proximity of a humanoid robot, the robot needs
awareness of its surrounding environment and the people in it, such that it can
perceive humans and the objects that can be involved in human-robot
interaction, anticipate human actions and potential chain of events in a
dynamic environment, and react appropriately to prevent hazardous situations.

Just as importantly, individuals need to be able to predict the robot’s behavior. A

humanoid robot is composed of a large number of moving parts compared to
other robots, which could affect predictability when the robot moves in ways
that are not quite human-like. Predictability can be facilitated with predictable
robot movements (for example, smooth trajectories with easily interpretable
goals) that are communicated by the robot – see the above sections for details
on communication types and behaviors. Maintaining predictability and
communication is especially critical during HRI, where changes in robot
motions may be required to dynamically accommodate hazard avoidance.

Just as with any other mobile manipulator, a humanoid’sphysical design is also

critical for safe HRI: lightweight components, soft materials, rounded edges, and
compliant joints (intrinsically or through control) can reduce the potential for
injury. Additional considerations in the control systems may be needed to
account for the safety hazards that could stem from the many moving parts on
the robot, the behavior of which a human may not be fully aware, and which
may result in unpredictable motions when the humanoid’s balance is affected
through HRI.

To remain mobile while carrying loads, a humanoid robot’s motors may also be
more powerful than those in traditional cobots. Maintaining safety in this case
may call for the implementation of robust safety systems, such as fail-safe

82 of 95
mechanisms, error detection, safe stop, and error recovery mechanisms that are
designed with the assumption that humans may be nearby. The presence of
humans may also need to be taken into account for the robot to safely respond
to power loss/fluctuation, localization errors, and uncertainty in dynamic
situational awareness.

While the most critical concerns relate to physical safety, psychological safety
may need further consideration. In particular, how safe a robot is may not
always directly correlate with how much humans trust a robot, which will affect
human-robot interactions as described next.

Defining a “Trustworthy” Humanoid Robot

If humans are expected to exist or move through environments in which

humanoid robots are active, a form of social contract is implied between the
humans and the robots’ manufacturers, integrators, and owners. Specifically,
humans are extending a degree of trust to the robots and the entities that they
represent; trust that the robots will not intentionally harm people, or indirectly
contribute to conditions that could lead to harm. To build a trustworthy
humanoid robot, survey respondents indicated that focus should be placed on
several key elements.

As described in Robot Appearance and Communication Methods, from the

start, a humanoid robot must communicate its abilities and limitations for
humans to adjust their expectations and avoid over- or under-trust. Throughout
individuals’ interactions with a humanoid robot, it must demonstrate consistent
reliability, capability, and predictability in performing its tasks. While clear
communication is already critical for safety, transparency reinforces trust. Given
the complexity of humanoid robots, having a robot convey and explain its
intentions, actions, and decision-making can be beneficial. Additionally, given
the large data gathering capacity of humanoid robots, transparency on how
user data is collected and used, as well as evidence of robust privacy protection
measures and ethical decision-making processes, would affect how much trust
individuals place in their interactions with humanoid robots.

83 of 95
 Minimum Safety-Specific Behaviors for Human
Environments

When asked what they would consider as minimum safety-specific behaviors

for a humanoid robot, survey respondents further emphasized the following
aspects. To ensure safety, humanoid robots must be equipped with reliable
emergency stop mechanisms, offering various levels of halting, from immediate
power removal to controlled stops, including moving back to a safe state. Stop
mechanisms should includeboth autonomous and human-activated ones,
provided that the latter is accessible to the humans interacting with the robot.
This may require careful consideration, given humanoid robots’ high mobility.

A physical emergency stop button installed on the body of a humanoid robot

would ensure it is always in the proximity of the robot, but requires that
attempting physical human-robot interaction be safe (and feasible for
individuals with different (dis)abilities) at the moment where an emergency
stop may need to be activated. Novel, accessible approaches to remote
emergency stop activation may be necessary. When humans are in the
proximity of a humanoid, they may also need the ability to reason about the
safest courses of action, especially in the case of a loss of balance. Humanoids
need to be equipped with robust collision detection, avoidance, and reaction
systems that can detect and react to impacts with humans and objects before,
during, and after they happen, while also managing the robot's stability and the
stability of anything it is handling.

Given humanoid robots modelled on the human body, whether front-facing

cameras mounted on the head are sufficient, or if additional sensing is required
for the robot to maintain 360-degree awareness, may need to be carefully
assessed. Ensuring robot movements are safe for interactions with humans is
paramount, requiring smooth, controlled motions with adjustable speed and
force limits. Clear, bi-directional communication is essential, with multimodal
signals conveying the robot's intentions and alerting nearby humans of any
issues or hazards, and mechanisms allowing the robot to perceive and interpret
human communication (for example, confirmation that the robot may proceed
with a task).

Internal fault detection and diagnostic capabilities enable timely maintenance

and, if necessary, safe shutdowns. Humans must retain the ability to easily
override the robot's actions, and any object handling must be performed with

84 of 95
secure grasping and releasing to prevent hazards. Rigorous testing and
validation of all safety features in realistic scenarios will be crucial to guarantee
safe operation.

Humanoid robot safety requirements can vary significantly depending on the

environment in which they operate, which will affect the kinds of interactions a
humanoid robot is likely to encounter and prioritize as part of its task. While
some hazards are intrinsic to the environment itself, others manifest through
the interactions between agents and the environment. Movable objects can
become dislodged and impact people, robots, and other environmental
features. People and robots moving around can unintentionally close routes of
ingress and egress.

Pinching and crushing hazards can manifest when movable objects are put
near immovable features. Changes in operational conditions (e.g., lighting,
surface textures, and clutter), some of which may be dynamically introduced by
human activities, can make it difficult–if not impossible–to safely and reliably
move through a given environment.

When balancing risks to humans, to itself, or objects and beings in the

environment, a common narrative is that a humanoid robot must operate
under the unwavering principle that human safety is paramount, superseding
all other concerns. However, when it comes to risk to itself or objects and other
beings in the environment, beyond what may be hardcoded, contextual
information and interactions with humans may influence priorities. For instance,
a human could specify what the robot should be careful with and what should
be prioritized over the integrity of the humanoid robot itself. To satisfy these
requirements when operating within a dynamic environment, whether and
how a robot should be equipped with real-time risk assessment and lowest-risk
planning capabilities would need to be examined.

Human Supervision for Safe HRI

In some situations, a human supervisor may assist a humanoid robot's

operation, which introduces a distinct mode of human-robot interaction
through teleoperation and remote communication between the supervisor and
the robot. In some cases, a human supervisor may also need to communicate

85 of 95
with nearby people through the intermediary of the robot. The human
supervisor's role will be fundamentally shaped by the level of environmental
uncertainty.

In low-uncertainty environments, characterized by predictable routines and

well-defined tasks, and a low probability of human presence disrupting robot
operation, would allow for a more passive supervisory approach. Here, the
supervisor would act primarily as a monitor and overseer, requiring timely
checks to ensure optimal performance and anticipate maintenance needs,
consistent monitoring to detect deviations from expected robot behavior and
environmental conditions, and quick intervention when unexpected events
occur.

In contrast, in high-uncertainty environments, where humans are likely to be

present, unexpected events are frequent, and the robot's tasks are complex or
novel, a human supervisor would be called for to assume a highly active,
"hands-on" role. This is a demanding role for a human supervisor, requiring
them to remotely maintain awareness of the robot’s dynamic surroundings,
make rapid diagnoses and decisions to mitigate risks, operate the robot,
potentially overriding the robot’s autonomous actions, and coordinate with
nearby humans.

For a system as complex as a humanoid robot, the design of the user interface
is critical to (i) adequately communicate the robot’s capabilities and
limitations, but most importantly, to (ii) ensure the supervisor is not
overloaded (mentally, physically, temporally, …) and that their involvement does
not introduce further hazards.

Conclusion and HRI-Related Recommendations

As outlined across this chapter, individuals expect human-like abilities from

humanoid robots; for some, this goes to the point of expecting that a robot can
recognize and respond to human emotions. Some users may desire humanoid
robots to engage in meaningful conversations and emotional connections.
These high expectations, going well beyond functional task completion, are
likely influenced by the human-like shape of a humanoid robot. As a result,
robot designers may need to carefully consider the design of a robot to
accurately convey its abilities to the humans it will interact with.

86 of 95
Humanoid robots also need the ability to communicate their intentions, state,
and awareness of the environment. A variety of signals may be considered in
communication implementation, involving various modalities including audio
(verbal or nonverbal), visual, or physical signals, whether used explicitly,
implicitly, or through external devices. However, the interactive capabilities of a
humanoid robot need to be adapted to the needs and abilities of the individuals
with whom it will interact, and the environment in which it will do so. Special
attention to accessibility is necessary when working with individuals with
disabilities or members of vulnerable populations.

When humanoid robots share space and interact with humans, functionality
and reliability are critical, as are the inclusion of robust safety mechanisms,
including human supervision for uncertain situations. In some scenarios, the
benefits of robotic assistance may need to be balanced with potential harms,
physical or emotional, that may occur during human-humanoid robot
interaction. As part of risk assessment, robot malfunctions ranging from minor
issues to severe physical or emotional harm will need to be considered. How to
define acceptable rates of humanoid robot malfunctions within different use
cases remains an open question.

Concerns relating to privacy, data collection, and transparency were only briefly
touched upon, but will affect the trust people place in their interactions with
humanoid robots. It is also worth reinforcing that the themes covered in this
chapter were those that the team considered the most critical. However,
additional concerns about interacting with humanoids may have been missed
and may surface as humanoid robots are being developed, tested, and
deployed.

While the standards remain unfinished, manufacturers should be aware of

intended standards and begin building capabilities into their robots (i.e., lights
for communication, shaping of motions to be faster/slower, smoother/more
robotic) such that standards can be met when they are implemented.
Organizing this could be another job for what the IEEE study group becomes in
the future.

88 of 95
 Report Summary: Building a Standards Framework for
Humanoid Robots – Classification, Stability, and
Human-Robot Interaction

Humanoid robots are approaching a tipping point in development, promising

general-purpose functionality across industrial, service, and public applications.
However, widespread deployment is limited by one overriding challenge: the
lack of standards designed for the unique risks and capabilities of humanoids.

A coordinated effort among Standards Development Organizations (SDOs) is

urgently needed to establish a structured framework. This framework should be
built on three interconnected pillars—Classification, Stability, and Human-Robot
Interaction (HRI), each informing the others to form a comprehensive standards
pathway.

Classification: The Foundation of All Standards

The first step in any standardization effort is a clear, agreed-upon classification

system. Current robotics standards were designed for fixed-base or statically
stable robots, not for humanoids that combine bipedal locomotion, dynamic
balance, and manipulative dexterity. Without a common taxonomy, regulators,
manufacturers, and end users cannot consistently determine which standards
apply or what performance expectations are realistic.

A multi-layered classification framework is recommended, encompassing:

• Physical Capabilities – locomotion type (bipedal, hybrid), dexterity level,

and sensory systems.
• Behavioral Complexity – degree of autonomy, manipulation skills, and
adaptive behaviors.
• Application Domains – industrial, healthcare, public service, and other
specialized use cases.
• Humanoid-Specific Traits – anthropomorphic resemblance, naturalistic
motion, and interaction modalities.

88 of 95
For SDOs, this classification system can serve as the “table of contents” for
future standards. It allows committees to map which standards are broadly
applicable (e.g., functional safety from ISO 13849) and which need humanoid-
specific extensions (e.g., balance safety, fall-response behaviors).

Stability: The Critical Bottleneck for Safety and

Performance

If classification is the foundation, then stability becomes the obstacle that must
be overcome for humanoids to operate effectively in shared human spaces.
Unlike wheeled or fixed robots, humanoids constantly deal with managed
instability; even powered-down robots can fall, which creates inherent hazards.

Key gaps identified in current standards include:

• No quantifiable stability metrics tailored for actively balancing robots.

Existing measures, such as those in industrial mobile robot standards, only
consider tip-over from payload shifts, not active balance control.
• No standardized test methods exist to evaluate stability performance
under realistic tasks, disturbances, or environmental variability.
• Functional safety models that do not account for dynamic balance—
current SIL and PL measures assume deterministic systems, whereas
humanoids require predictive, probabilistic risk modeling.

A two-phase standards roadmap is recommended:

• Performance Standards (ASTM, IEEE) – Develop stability metrics (e.g.,

margin of stability, capture point, predictive stability regions) and task-
based test methods (walking on uneven terrain, carrying loads, recovering
from pushes).
• Safety Standards (ISO, IEC) – Translate those performance metrics into
minimum safety thresholds, incorporating predictive control
requirements, fall-handling behaviors, and residual risk limits.

For SDO members, stability testing and safety validation must be treated as
intertwined efforts. ASTM and IEEE can lead the development of repeatable test
methods and quantifiable metrics, while ISO and IEC integrate these into
regulatory-grade safety standards.

89 of 95
 Human-Robot Interaction: Managing Risk and
Perception

Humanoids occupy a unique space in human-robot interaction (HRI) because of

their Anthropomorphic form. People naturally project human-like expectations
onto them, resulting in two major risks: overtrust and psychosocial discomfort.

From a safety perspective, humanoids introduce indirect risks that are not
covered by traditional robot safety standards. Automating a process with
humanoids can alter workflow pacing and repetition, thereby increasing
musculoskeletal and cognitive risks for human workers in both pre- and post-
automation tasks.

From a perception standpoint, technically safe motions can still feel unsafe; fast
limb swings, sudden steps, or a robot standing too close can cause discomfort,
especially in public environments where bystanders are untrained.

Future HRI standards should therefore address:

• Collaborative Task Safety Thresholds – Extending ISO/TS 15066 principles

for humanoid mobility and manipulation, including safe-fall zones and
controlled stance modes when humans are nearby.
• Interpretable Behavior and Body Language – Defining motion guidelines
to signal intent.
• User Training and Expectation Management – Standardizing how
capabilities are communicated to avoid overestimation of performance.

Classification and stability standards directly inform these HRI standards. For
example, a humanoid classified for public use would require higher stability
thresholds, stricter fall-response requirements, and interaction-specific body
language guidelines than one confined to a closed industrial setting.

Constructing the Standards Framework

The interdependence of classification, stability, and HRI highlights the need for
a coordinated, multi-SDO approach rather than piecemeal adaptations of
existing standards. A suggested pathway includes:

• Classification as the Organizing Principle

90 of 95
 • Adopt a shared taxonomy across SDOs to define humanoid types, guiding
which existing standards can apply and where gaps remain.
• Parallel Development of Stability Metrics and Test Methods
• ASTM and IEEE can lead to task-based performance test method creation
(e.g., walking, manipulation under disturbances).
• IEEE can standardize stability metrics for predictive and instantaneous
balance assessment.
• Integration into Safety Standards
• ISO and IEC can develop application-specific safety thresholds,
incorporating test methods and metrics into regulatory safety validation.
• HRI and Perception-Based Standards
• Build on classification and stability results to create interaction guidelines
addressing both physical and psychosocial safety.
• Centralized Coordination
• Establish a joint working group spanning ISO, IEEE, and ASTM to ensure
alignment. A shared roadmap would prioritize stability-related standards
first, as they are the primary barrier to safe adoption.

Toward Trustworthy Humanoids

The three pillars are not isolated; they are mutually reinforcing. Classification
clarifies what type of humanoid is being evaluated, stability standards prove
that it can function safely, and HRI standards ensure it does so in ways that
humans find acceptable and trustworthy.

Until these elements are developed in tandem, humanoids will remain limited
to controlled environments and pilot programs. But with a coordinated
standards effort, SDOs have the opportunity to build the framework that will
make humanoids reliable, certifiable, and ultimately, deployable in the diverse
human spaces for which they are designed.

91 of 95
Special Thanks To The Following Contributors Who
Provided Insights Into The Development Of This
Report

Francisco Andrade-Chavez Tanya Grinberg Aleshin Nikita

Chioma Aso William Harrison Mike Oitzman

Fahmi Badreddine Matthew Howard Rob Paolozzi

Christian Becker-Asano Michael Jensen Jim Pippine

Jan Butz Salim Karaer Deb Prince

Marie Charbonneau Sagar Kasrung Maximo Roa

Francisco Chavez Andra Keay Ted Rozier

Vincent Clerc Niamh Kingsley Kamel Saidi

Parker Conroy Jacques Lemire Kartik Sachdev

Alvaro Costa Shang-Ching Liu Joan-Wilhelm Scharze

Maharshi Digumarthi Pablo Lopez Garcia Werner Varro

Claude Dismoor Matthew Marge Tom Vardon

Nikolai Ensslen Riccardo Mariani Dieter Volpert

Carlos Felix Syrine Mansour Axel Wechsler

Toni Garcia Rodriguez Carlos Mastalli Ruoshi Wen

Animesh Garg Bill Mather Yuke Zhu

Julien Ghaye Jonathan McCormick Zhijie Zhao

Mona Ghassemian Nicholas Milton

92 of 95
Bibliography (Extracted from
Footnotes)
1. In this document, ‘self-balancing’ and ‘actively-balancing’ are used as
interchangeable terms when referring to a robot. They both describe a robot
that only remains upright through active, powered balance.
2. ISO 12100:2010 Safety of machinery — General principles for design — Risk
assessment and risk reduction
3. IEC 60204-1 (Consolidated version 2016+AMD1:2021) Safety of machinery -
Electrical equipment of machines - Part 1: General requirements
4. ISO 10218-1:2025 Robotics — Safety requirements Part 1: Industrial robots
5. ISO 10218-2:2025 Robotics — Safety requirements Part 2: Industrial robot
applications and robot cells
6. ISO 13482:2014 Robots and robotic devices — Safety requirements for
personal care robots
7. ISO 13850:2015 Safety of machinery — Emergency stop function — Principles
for design
8. ISO 13855:2024 Safety of machinery — Positioning of safeguards with
respect to the approach of the human body
9. IEC 61508 Functional safety of electrical/electronic/programmable electronic
safety-related systems - (Part 1 to 7)
10. ISO 13849-1:2023 Safety of machinery — Safety-related parts of control
systems Part 1: General principles for design, and ISO 13849-2:2012 Safety of
machinery — Safety-related parts of control systems Part 2: Validation
11. ANSI/RIA R15.08-1-2020 Industrial Mobile Robots - Safety Requirements -
Part 1: Requirements for the Industrial Mobile Robot
12. ANSI/ITSDF B56.5-2024 Safety Standard for Driverless, Automatic Guided
Industrial Vehicles and Automated Functions of Manned Industrial Vehicles
13. ISO/TR 20218-1:2018 Robotics — Safety design for industrial robot systems
Part 1: End-effectors
14. IEC 63310:2025 Functional performance criteria for AAL robots used in
connected home environment
15. ISO 7176-1:2014 Wheelchairs Part 1: Determination of static stability
16. ISO 7176-2:2017 Wheelchairs Part 2: Determination of dynamic stability of
electrically powered wheelchairs
17. ISO/TR 4448-1:2024(en) Intelligent transport systems — Public-area mobile
robots (PMR) — Part 1: Overview of paradigm
18. See full details and options for PFHd values in 5.3.3 of ISO 10218-1:2025
19. ISO 21448:2022 Road vehicles — Safety of the intended functionality
20. ISO 26262 - Road vehicles — Functional safety (parts 1 to 10)
21. ISO/AWI 25785-1 - Safety requirements for industrial mobile robots with
actively controlled stability (legged, wheeled, or other forms of locomotion) -
Part 1: Robots
22. ISO 9283:1998 Manipulating industrial robots — Performance criteria and
related test methods
23. ISO 18646 Robotics — Performance criteria and related test methods for
service robots (Part 1 to 8)
24. ISO/DIS 18646-5 Robotics — Performance criteria and related test methods
for service robots — Part 5: Locomotion for legged robots
25. F. Aller, D. Pinto-Fernandez, D. Torricelli, J. L. Pons and K. Mombaur, "From
the State of the Art of Assessment Metrics Toward Novel Concepts for
Humanoid Robot Locomotion Benchmarking," in IEEE Robotics and
Automation Letters, vol. 5, no. 2, pp. 914-920, April 2020
26. Ramuzat N, Stasse O, Boria S. Benchmarking Whole-Body Controllers on
the TALOS Humanoid Robot. Front Robot AI. 2022 Mar 4;9:826491.
27. C. Sferrazza, D.-M. Huang, X. Lin, Y. Lee, and P. Abbeel, “HumanoidBench:
Simulated Humanoid Benchmark for Whole-Body Locomotion and
Manipulation.” Robotics: Science and Systems Conference 2024
28. J. Baltes and S. Saeedvand, "Rock Climbing Benchmark for Humanoid
Robots," 2022 International Conference on Advanced Robotics and Intelligent
Systems (ARIS), Taipei, Taiwan, 2022, pp. 1-4,
29. Stasse O, Giraud-Esclasse K, Brousse E, Naveau M, Régnier R, Avrin G,
Souères P. Benchmarking the HRP-2 Humanoid Robot During Locomotion.
Front Robot AI. 2018 Nov 8;5:122.
30. “Eurobench project.” [Online]. Available: https://eurobench2020.eu/
31. W. Thibault, F. J. A. Chavez and K. Mombaur, "A Standardized Benchmark
for Humanoid Whole-Body Manipulation," 2022 IEEE-RAS 21st International
Conference on Humanoid Robots (Humanoids), Ginowan, Japan, 2022, pp.
608-615

94 of 95
 32. V. Lippi, T. Mergner, T. Seel and C. Maurer, "COMTEST Project: A Complete
Modular Test Stand for Human and Humanoid Posture Control and Balance,"
2019 IEEE-RAS 19th International Conference on Humanoid Robots
(Humanoids), Toronto, ON, Canada, 2019, pp. 630-635
33. Peng, W. Z., Song, H., and Kim, J. H. (March 12, 2021). "Stability Region-Based
Analysis of Walking and Push Recovery Control." ASME. J. Mechanisms
Robotics. June 2021; 13(3): 031005.
34. J. A. Castano, E. M. Espuela, J. R. Rojas, E. M. Hoffman, and C. Zhou,
“Benchmarking standing stability for bipedal robots,” IEEE Access, p. 1, Jan.
2025, doi: 10.1109/access.2025.3529191.
35. C. Curtze, T. J. W. Buurke, and C. McCrum, “Notes on the margin of
stability,” Journal of Biomechanics, vol. 166, p. 112045, Mar. 2024, doi: 10.1016/
j.jbiomech.2024.112045.
36. T. Mergner and V. Lippi, “Posture Control—Human-Inspired Approaches for
Humanoid Robot benchmarking: conceptualizing tests, protocols and
analyses,” Frontiers in Neurorobotics, vol. 12, May 2018, doi: 10.3389/fnbot.
2018.00021.
37. ISO 13482:2014 “Robots and robotic devices — Safety requirements for
personal care robots” has some relevant references for human-robot
interactions but it is specifically dedicated to service applications and under
deep revision at the time of publication of this Report, so omitted as an option.
38. This Technical Specification is soon to be withdrawn as it is incorporated in
the revised series ISO 10218-1 (and 2):2025 “Robotics — Safety requirements -
Part 1 (and 2)”.
39. S. Kajita et al., "Impact acceleration of falling humanoid robot with an
airbag," 2016 IEEE-RAS 16th International Conference on Humanoid Robots
(Humanoids), Cancun, Mexico, 2016, pp. 637-643,
40. Elo, S. and Kyngäs, H. (2008), The qualitative content analysis process.
Journal of Advanced Nursing, 62: 107-115. https://doi.org/10.1111/j.
1365-2648.2007.04569.x https://doi.org/10.1111/j.1365-2648.2007.04569.x
41. Greta Hilburn, UX Researcher and Designer, US Defense Acquisition
University

95 of 95
View publication stats