I. Robotics and Perception Group, UZH

I spent a year in the RPG lab starting on June 2016 as a visiting scientist. I supervised multiple student projects on themes of trajectory optimization and modelling of quadrotors using the drake state estimation and control library.

Fast trajectory optimization of quadrotors with cable-suspended payloads.

We developed an algorithm for fast trajectory optimization for agile-motions that was inspired by state of art methods for planning through contact in legged robots. The resulting  paper was finalist for Best Student paper at RSS 2017.

Publication:

Foehn, P., Falanga, D., Kuppuswamy, N., Tedrake, R., & Scaramuzza, D. (2017). Fast trajectory optimization for agile quadrotor maneuvers with a cable-suspended payload. (Link)

II. Dynamic Interaction Control Lab, IIT

From July 2014- June 2016, I was a member of the Dynamic Interaction Control Group of the iCub Facility of Istituto Italiano di Tecnologia (Italian Institute of Technology). Our group, under the direction of Francesco Nori has focussed on developing state-of-art methods for humanoid robots to control their interaction forces; the result is that such robots can exploit their dynamics and suitably perform task in the complex environments such as homes.

My research encompassed the following 3 directions :

a. Probabilistic State Estimation for Articulated Robots

A valuable component necessary for the control of interaction forces (force / torque control strategies) is an effective approach for whole-body state estimation. In this research, I have explored the following projects along with my colleagues.

ii. Whole-body state estimation for humanoids

We developed a framework for whole-body state estimation that combined multiple distributed sensory modalities (accelerometers, gyroscopes, F/T sensors, joint encoders and contact-pressure tactile sensors) in order to probabilistically estimate the state of the robot (configuration space & external forces) exploiting the natural sparsity of the multi-body dynamics.

Sample publication :

Nori, Francesco; Kuppuswamy, Naveen; Traversaro, Silvio, “Simultaneous state and dynamics estimation in articulated structures,” in Intelligent Robots and Systems (IROS), 2015 IEEE/RSJ International Conference on , vol., no., pp.3380-3386, Sept. 28 2015-Oct. 2 2015
doi: 10.1109/IROS.2015.7353848
IEEE Explore page

ii. Foot state estimation for legged robots

A special case of probablistic estimation for legged robots is in accurately estimating the foot state (configuration / external forces) by incorporating multimodal sensing and rigid body assumptions on individual feet. Applications include accurate prediction of balance metrics (Foot rotation indicator) due to unforeseen incidents like loss of balance.

Sample Publication :

Eljaik, Jorhabib; Kuppuswamy, Naveen; Nori, Francesco, “Multimodal sensor fusion for foot state estimation in bipedal robots using the Extended Kalman Filter,” in Intelligent Robots and Systems (IROS), 2015 IEEE/RSJ International Conference on , vol., no., pp.2698-2704, Sept. 28 2015-Oct. 2 2015
doi: 10.1109/IROS.2015.7353746
IEEE Explore Page 

b. Human Dynamics estimation from motion capture

Using the methods developed for whole body state estimation for humanoids, we are currently developing strategies for dynamics estimation on human subjects using motion capture and external force sensing. We are progressively moving towards developing a wearable system for dynamics estimation on human subjects performing various tasks.

Sample Publication

C Latella, N Kuppuswamy, F Romano, S Traversaro, F Nori “Whole-Body Human Inverse Dynamics with Distributed Micro-Accelerometers, Gyros and Force Sensing“, Sensors 16 (5), 727.

c. Optimal Impedance for walking on uneven terrains

In walking on uneven terrain, a key challenge is to cope with an unexpected orientation at the moment of contact. Impedance control can play a crucial role in mitigating the instability caused by the foot contact with the terrain. We analytically compute an optimal impedance for stabilising the robot in such situations by maximising a cost function based on foot center-of-pressure after full foot placement establishment.

Sample Publication

Romano, F., Kuppuswamy, N., Ciocca, M., Traversaro, S., & Nori, F. (2016, November). Stable bipedal foot planting on uneven terrain through optimal ankle impedance. In 2016 IEEE-RAS 16th International Conference on Humanoid Robots (Humanoids) (pp. 146-151). IEEE. (IEEE / download link)

d. Self-calibration of joint offsets for multi-link systems

Humanoids and other multi-link robots often utilize relative encoders and often require time-consuming self-calibration procedures to “zero” them accurately; such calibration may be infeasible when deploying in the field. This project explored an approach inspired by our whole-body estimation framework for incorporating multiple accelerometer measurements within a self-calibration algorithm (IEEE / download link).

Sample Publication

Guedelha, N., Kuppuswamy, N., Traversaro, S., & Nori, F. (2016, November). Self-calibration of joint offsets for humanoid robots using accelerometer measurements. In 2016 IEEE-RAS 16th International Conference on Humanoid Robots (Humanoids) (pp. 1233-1238). IEEE.

III. AI Lab, University of Zurich

I was at the AI Lab of the University of Zurich between February 2009 and January 2014. My research was supported by three EU Projects. My thesis tied together the various projects I worked on in that time.

a. PhD Thesis: Exploiting Reduced Dimensionality in Design and Control of embodied systems

High dimensionality poses several questions in our current understanding of the mechanisms of control and learning in nature. The key problem lies in understanding the principles by which nature copes with high-dimensional control and optimisation. Solving it is not only essential for a theory of motor control in embodied systems; it could also yield novel methods for the design and control of high-dimensional biomimetic artificial systems. In this thesis a possible solution to this problem is proposed in the form of a design principle of exploiting reduced dimensionality. Reduced dimensionality is defined as the property of a system by which dimensionality reduction of its dynamics and behaviour is facilitated. This notion is formalised mathematically by a proposed control-theoretic framework of reduced dimensionality analysis. The framework requires the definition of a quality factor for the feasibility of the reduced dimensional model for control and the choice of an appropriate model reduction algorithm. Using this framework, the neuroscientific hypotheses of optimal motor control, developmental skill acquisition, and muscle synergies are analysed synthetically. The framework is used for a systematic study of the different factors affecting reduced dimensionality , i.e. (i) natural dynamics, (ii) output (task-space relevance) and, (iii) input (modularisation of the control). Case studies further examine reduced dimensionality in the viewpoint of (iv) Dynamical Movement Primitives for robotics and, (v) dimensional change accompanying development of motor skills. Both theoretical analyses and empirical demonstrations using simulations are presented for each of the studies. The results indicate that reduced dimensionality can be effectively exploited as a design principle for embodied systems. They have implications for both biological theories of motor control and development and for the design and control of high-dimensional artificial systems.

Download link

b. Amarsi Project (FET FP7)

AMARSi (Adaptive Modular Architecture for Rich Motor Skills) aimed at a qualitative jump in robotic motor skills towards biological richness. Rich motor skills in robots will have a tremendous impact on our society. Dexterous and skillful motion in robots will make them more suitable for a large number of tasks. The compliant and natural movements will make them blend into everyday routines, safe and psychologically acceptable.

I was a member of the AMARSi project between September 2012 and September 2013 and worked on two lines of research :

i. Muscle synergies, dimensionality reduction and minimum dimensional control

Understanding of Muscle Synergies from the perspective of minimum dimensional control. I developed a technique of Trajectory Specific Dimensionality Analysis (TSDA) to analyse the relationship between trajectories and dimensionality of a dynamical system that can generate them. The approach is based on tools from nonlinear model order reduction.

Sample Publication :

Kuppuswamy, N., & Harris, C. M. (2014). Do muscle synergies reduce the dimensionality of behavior? Frontiers in Computational Neuroscience, 8, 63. http://doi.org/10.3389/fncom.2014.00063
Frontiers paper page.

ii. Motor primitives in robots

Using methods for model order reduction, I explored the unsupervised learning of reduced dimensional control for compliant redundant robot systems. In particular, I used the Pendulum Robot as a test platform.

Sample Publication :

N. Kuppuswamy, H. G. Marques, and H. Hauser. Synthesising a motor-primitive inspired control architecture for redundant compliant robots. In Tom Ziemke, Christian Balkenius, and John Hallam, editors, From Animals to Animats 12, volume 7426 of Lecture Notes in Computer Science, pages 96–105. Springer Berlin Heidelberg, 2012.

Lecture Notes in Computer Science link.

c. RobotDoC (Marie Curie ITN)

The RobotDoC (Robotics for Development of Cognition) is a Marie Curie Initial Training Network that I was a part of from September 2009 – September 2012. The RobotDoC Collegium was a multi-national doctoral training network for the interdisciplinary training on developmental cognitive robotics. and the network comprised of 17 Marie Curie Fellows including me. While some of the research outcomes have overlapped with my research in AMARSi (above), an overview of my activities can be seen in the the video below:

d. Octopus Project (FET FP7)

The octopus project was aimed at investigating and understanding the principles that give rise to the octopus sensory-motor capabilities and incorporating them in new design approaches and ICT and robotics technologies to build an embodied artefact (i.e. a novel octopus inspired soft robot system), based broadly on the anatomy of the 8-arm body of an octopus, and with similar performance in water, in terms of dexterity, speed, control, flexibility, and applicability.

I was a member of the Octopus project between February 2009 and September 2009. I developed the low level control hardware for the tendon driven octopus-inspired soft robot arm below: The arms were made using silicone and incorporated upto 16 different tendons.

Sample Publication :

Nakajima, K.; Tao Li; Kuppuswamy, N.; Pfeifer, R., “Harnessing the dynamics of a soft body with “timing”: Octopus inspired control via recurrent neural networks,” in Advanced Robotics (ICAR), 2011 15th International Conference on , vol., no., pp.277-284, 20-23 June 2011. doi: 10.1109/ICAR.2011.6088590

IEEE Explore Page

I also worked on modelling the behaviour of the soft continuum arm. I co-developed a curvature space linear dynamical model and performed some trials on identifying such a model. The objective was to then develop simple linear curvature space low level controllers which could then be combined for higher level cartesian task objectives.

e. Collaborations with Oliver Wolf

During my stay at the AI Lab, I also collaborated with Oliver Wolf, an artist-in-lab resident. I provided technical assistance for some of the art installations that he developed. The Geostationary Space travel installation was one such example.

IV. Yujin Robot Co. Ltd.

Yujin Robot is one of South Korea’s leading service robot companies and is based in Seoul. I worked there from August 2007-August 2008. I worked on two key projects: (i) Developing HRI applications using third party computer vision tools. (ii) Developing a preliminary version of service robot middleware – a precursor towards executing third party HRI services on a variety of robot platforms.

V. RIT Lab, KAIST

• Omnidirectional Football robot – Design and Control
• Episodic Memory for Cognitive Agents
• Ubiquitous Robotics

(More information available on request)