EE

Duality AI - Space Station Hackathon

1. Overview

Duality AI is excited to present the Space Station Hackathon, a challenge designed

to test cutting-edge AI training techniques to train a model for Object Detection in
a space station environment. Participants will train a model using a synthetic
dataset from Duality AI’s digital twin simulation platform- Falcon. Each team will
work to achieve the most accurate and precise model by adjusting training
parameters and processes. Along the way, participants will learn how Duality AI
uses high-quality synthetic data to train AI models in unseen or difficult-to-access
environments, such as a space station. Through industry-proven techniques and
tools, we can train an AI model in tasks such as object detection using only
synthetic data. Through this competition, participants will develop novel AI training
skills, enhance their portfolio, build connections with Duality AI and other
participants, and have the opportunity to win prizes and recognition!

Objectives

●​ Train a robust object detection model using the provided synthetic dataset
to accurately detect and classify space station objects — a critical
capability for ensuring operational safety.
●​ Evaluate model performance across challenging scenarios including varied
lighting, object angles, and occlusions.

1
 ●​ Benchmark and optimize your model to improve accuracy, generalizability,
and efficiency for real-world deployment scenarios.
●​ (For bonus points) Create an application that uses this detection model,
including how to keep the model up to date using Falcon.

Importance of Digital Twins in Space Station Simulation

Digital twins are highly realistic virtual representations of real-world systems. This
challenge uses data from a digital twin space station environment generated
within Falcon. Digital twins, such as the space station, allow students,
researchers, and AI enthusiasts to train a model in an environment that they
otherwise might not be able to in the physical world - whether because it is
difficult to access, not yet in existence, or limited in its availability. Using synthetic
data from Falcon, users can:

●​ Simulate real-world conditions (e.g., lighting, occlusions).

●​ Train AI models without relying on real-world data collection.
●​ Enable reproducible and scalable experimentation.

Hackathon Structure

Data Overview
●​ Participants will work with a dataset generated from FalconEditor of a
simulated space station scene featuring:
○​ Varying lighting conditions
○​ Occlusions and object overlap
○​ Diverse camera angles and distances
●​ Target Object Categories:
○​ Toolbox
○​ Oxygen Tank
○​ Fire Extinguisher

2
Participants will process synthetic data from the simulation, train an AI model to
classify these objects, and optimize detection accuracy under realistic constraints.

i. Hackathon Tasks

To maximize efficiency, we recommend dividing responsibilities based on these

roles. Proper delegation will help streamline the workflow and ensure high-quality
results across the technical and presentation components.

1.​ AI Engineering
●​ Train and fine-tune the YOLOv8 object detection model using the
generated dataset.
●​ Evaluate model performance.
●​ Use optimization techniques to improve accuracy and reduce inference
time.

2.​ Documentation & Presenting Final Team Results

●​ Document the workflow, including:
○​ dataset generation
○​ model training
○​ evaluation metrics.
●​ Prepare a structured report and final presentation showcasing findings,
results, and insights.
●​ Create visualizations such as confusion matrices and performance
graphs to highlight model behavior.

3.​ (Bonus) Create an App

●​ Create a computer or phone app that uses the model
●​ Describe the plan for keeping the model up-to-date using Falcon

3
ii. Key Deliverables

At the end of the hackathon, teams must submit:

1.​ Trained Object Detection Model

a.​ Fully trained YOLOv8 model that detects and classifies all three
object categories.
b.​ The package must include model weights, scripts, and config files.

2.​ Performance Evaluation & Analysis Report, including but is not limited to:
a.​ mAP Scores (@0.5 IoU) to evaluate model accuracy.
b.​ Confusion Matrix to visualize class-wise performance.
c.​ Failure Case Analysis, highlighting misclassifications and possible
improvements.

iii. Prizes and Rewards

We are excited to offer the following prizes and recognition to participants in the
Duality Track - Space Station Hackathon:

●​ First Place: $200

●​ Second Place: $100
●​ Top 10 Teams: Participation Certificate
These prizes are awarded based on the performance of your object detection
model, with a focus on accuracy and innovation. We encourage all teams to strive
for excellence and showcase their skills in AI model training and optimization.

iv. Judging Criteria

Teams will be evaluated on a 100-point scale, based on the following criteria:

4
a. Model Performance
[email protected] (Mean Average Precision at IoU=50%) …………..…...……80 Points
○​ Measures the accuracy of object detection and classification.
○​ Higher mAP scores indicate better model performance in detecting
and classifying objects correctly.

b. Performance Report Clarity

Structured Findings & Detailed Reporting……………………………….…20 Points
○​ Well-organized documentation of the methodology, challenges, and
solutions.
○​ Clearly outlined steps, including dataset creation, model training, and
evaluation.

c. (Bonus) Use Case Proposal

Proposed Application …………………………………………………(Bonus) 15 Points*

*Partial completion can still give you partial points. You can only score a
maximum of 100 points for this Hackathon challenge, but you can earn 15 bonus
points that will be added to your final score UP TO 100 total points.

○​ Create an application that uses this model

○​ Include how Falcon can be used to keep the model up-to-date with
the real world use

Teams must balance technical performance with clear and professional

documentation to maximize their score.

2. Task Instructions

5
i. AI Engineering:

This section will walk you through:

➔​ Setting up your Falcon account
➔​ Downloading and using the dataset
➔​ Preparing your training environment
➔​ Understanding the training workflow (Exercise 3)
➔​ Establishing baseline performance with YOLOv8

1.​ Create a Falcon Account:

a.​ Visit Falcon and sign up for an account if you don’t have one
b.​ Once registered, log in to access datasets, exercises, and tools.

2.​ Download the Dataset

a.​ Download the dataset to your local machine from this page
i.​ The page also contains useful links
ii.​ The dataset includes:
■​ Pre-collected images and YOLO- compatible labels
a.​ Separated into Train, Val, and Test folders
■​ Sample Training and Prediction scripts
■​ Configuration files (config.yaml)

3.​ Set Up the Training Environment

a.​ Make sure you have Anaconda installed, and open an Anaconda
Prompt window
b.​ Navigate to the ENV_SETUP sub folder in the Hackthon_Dataset
folder
c.​ Run the setup_env.bat file in the Anaconda Prompt window
i.​ This will set up an environment called “EDU”, containing all the
dependencies required to run the training and prediction
scripts.

6
 NOTE- Mac/Linux users, Create a setup_env.sh script with equivalent
commands

4.​ Understand the Training Workflow using Exercise 3

Falcon provides a guided walkthrough, explaining:
a.​ How the YOLOv8 training pipeline is structured
b.​ What each training script does
c.​ How to evaluate and visualize results

5.​ Train the YOLOv8 Model on the Sample Dataset

Once the EDU environment is ready and the dataset is in place:
a.​ Open an anaconda command prompt or a terminal
b.​ Navigate to the training and prediction scripts directory
c.​ Activate the environment by typing ‘conda activate EDU’ in the terminal.
d.​ Run the training command: ‘python train.py‘
This will begin training your model and save logs + checkpoints to the runs/
directory.

6.​ Establish Benchmark Results with YOLOv8 Model on the sample dataset
After training is completed, evaluate your model’s performance by running the
prediction script (predict.py) in the same command prompt window. This tests its
performance on real-world test images and gives you the following:
a.​ [email protected]
b.​ Precision and Recall
c.​ Confusion Matrix
d.​ Predictions

Use the results as your benchmark, so that later, when you train with newer model
settings, you can:
●​ Compare performance
●​ Track improvements
●​ Identify where the model struggled (e.g., occlusion or lighting)

7
ii. Documentation & Presentation:

Ensure that your team’s work is:

✔ Clear and understandable

✔ Reproducible by others
✔ Professional and impactful for judges

1.​ Keep it Structured & Organized using this General Structure

a.​ Title & Summary: Clearly state the purpose of the document.
b.​ Step-by-Step Instructions: Use numbered lists or bullet points.
c.​ Diagrams & Visuals: Use flowcharts, tables, and screenshots from
the training process and the visuals from the runs/ folder (created
after training and prediction is complete)
i.​ Graphs & Charts: Show training loss, accuracy trends, and
comparisons.
ii.​ Screenshots: Use images from the runs/ folder after training
iii.​ Confusion Matrix: Highlight model performance across
different classes.
iv.​ Before & After Images: Show examples of correct vs.
misclassified objects.

2.​ Document Everything, But Keep it Concise

a.​ Record major steps like dataset creation, training process, and
evaluation.
b.​ Avoid overly technical language — aim for clarity.
c.​ Use clear, plain language—assume the reader is new to the project.

3.​ Example Entry:

a.​ Task: Model Training on Dataset
b.​ Initial mAP Score: 42%

8
 c.​ Issue Faced: Low recall for "Oxygen Tank" class due to
occlusion.
d.​ Solution: Augmented dataset with occlusion examples → New
mAP Score: 55%

4.​ Documenting Failure Cases and Solutions

a.​ Include failure case images to illustrate what went wrong and how it
was fixed.

5.​ Report Format (8 Pages Max)

Your Report should be concise, structured, and visually engaging. Use the following
storytelling approach:

Problem → Fix→ Results → Challenges → Future Work

Page.No Section Description

1 Title Team name, project name, brief

tagline.

2 Methodology Steps taken while training the

model, and fine-tuned results.

3-4 Results & Performance mAP scores, confusion matrix,

Metrics accuracy comparisons.

5-6 Challenges & Solutions Key obstacles and how they

were resolved.
*See below for examples

7 Conclusion & Future Work Final thoughts, and potential

improvements.

9
iii. (Bonus) Use Case Proposal:
This is an optional bonus task that can replace any of the required points. It can push the score
up to, but not above, 100 total points.

1.​ Create an application that uses this object detection model and runs
either on your computer or phone
a.​ Show how this model could be used to address a real-world problem
or need.
b.​ Create a video of the application, or share the application files in your
submission
2.​ Include how Falcon can maintain the model
a.​ As part of the application, describe how Falcon could continuously be
used to update the model as the real-world problem changes.
i.​ For example: If the detected object changes its look or if a new
object gets introduced to the environment that causes
confusion for the model, how can Falcon be used to address
the new changes?

3. Submission and Final Steps Instructions

i. Necessary Submission Files:

●​ A single, Final Packaged Folder that includes all necessary files:

​Model training and inference scripts (train.py, predict.py)
​YOLO configuration files

10
 ​runs/directory (containing training logs, outputs, and performance
visualizations)
​Any additional assets or scripts required to test your model

●​ A well-structured Hackathon Report (PDF or DOCX) that covers the

following:
​Methodology: Your training approach and setup
​Challenges & Solutions: Issues faced and how you overcame them
​Optimizations: Techniques used to improve model performance
​Performance Evaluation: [email protected] score, Confusion matrix, Failure
case analysis and observations.

●​ A README.md or README.txt that provides:

​Step-by-step instructions to run and test your model
​How to reproduce your final results
​Any environment or dependency requirements
​Notes on expected outputs and how to interpret them

●​ (BONUS) A Use Case Application that covers the following:

​A .pdf or .docs file that describes what your application does, how
you created the application, and a proposed plan for updating the
model as the situation changes over time
​A video of the application working, or the application execution
files/links

Note: You are welcome to use your own object detection models and custom
training scripts or notebooks. However, you must train your model exclusively on
the dataset provided for this challenge.

11
Important: Using any of the designated testing images for training purposes is
strictly prohibited and will result in disqualification. Please ensure a clear
separation between training, validation, and test sets throughout your workflow.

ii. Upload Instructions:

1.​ Ensure your submission folder contains all of the above

2.​ Compress everything into a .zip file.
3.​ Upload the zipped folder to a private GitHub repository of your own.
4.​ Complete the submission form (Here) by:
a.​ Reporting your team’s final score
b.​ Providing the GitHub repository link
5.​ Finally, add the following reviewers as collaborators:
a.​ Rishikesh Jadhav
i.​ GitHub Username - Rishikesh-Jadhav
b.​ Syed Muhammad Maaz
i.​ GitHub Username - Maazsyedm
c.​ Rebekah Bogdanoff
i.​ GitHub Username - rebekah-bogdanoff

iii. After Submission:

1.​ Sharing Results & Feedback

a.​ After submission, teams can showcase their models and insights.
b.​ Feedback from judges will be provided after evaluation.
c.​ A selected few teams will be invited for a post-hackathon discussion.

2.​ Future Opportunities & Improvements

a.​ Continue exploring advanced topics such as:
i.​ Self-supervised learning
ii.​ Domain adaptation

12
 iii.​ Multi-view detection
b.​ Consider using Falcon to generate your own synthetic data
augmentation for better training.

3.​ Stay connected with us via Discord for:

a.​ Future challenges
b.​ Internship and apprenticeship opportunities
c.​ Community events and AI workshops

NOTE: If you face any issues along the process join our discord channel linked
above where we will be providing support to you for the hackathon.

4. Common Issues and Troubleshooting

FAQ’s
1.​ Will setup_env.bat Work on Both Windows & Mac?
○​ No, The setup_env.bat script is designed primarily for Windows
environments.
○​ Alternatively, for Mac/Linux use create a setup_env.sh shell script
equivalent that installs all required packages.
2.​ My training process is too slow. What can I do?
○​ Here are a few tips to improve training speed:
i.​ Reduce the batch size in your training configuration.
ii.​ Close any unused applications or background processes to
free up system resources.
iii.​ If using GPU, monitor GPU usage with tools like nvidia-smi.
3.​ How should I manage data transfer between team members or for
submission?
○​ Why backup your data:
i.​ Share results between teammates

13
 ii.​ Backup project files and checkpoints
○​ We recommend using cloud storage platforms such as Google Drive,
Dropbox, OneDrive, git
4.​ Where can I find the training and prediction scripts?
○​ Located inside the dataset folder: (train.py, predict.py)
5.​ What if my model is not improving?
○​ For additional tips to improve performance refer Here

Support and Communication

1.​ Join the official Duality Community Discord Server for:
a.​ Real-time help
b.​ Announcements
c.​ Live Q&A with organizers
d.​ Invite Link Here

5. Glossary and Metric Benchmarks

Term Definition

Digital Twin A virtual replica of a real-world object

or system.

Multi-Class Object Detection Identifying and classifying multiple

objects in an image/video.

mAP (Mean Average Precision) Primary metric for evaluating object

detection models.

IoU (Intersection over Union) Measures how well a predicted

bounding box overlaps with the ground
truth.

Bounding Box A rectangle defining an object's

position in an image.

14
NMS (Non-Maximum Suppression) A technique to remove duplicate
detections of the same object.

Class Imbalance When some object categories have

significantly fewer samples than
others.

Ground Truth Manually labeled data specifying

correct object locations and classes.

mAP (Mean Average Precision @ 0.5 Measures how well the model detects
IoU) objects (avg. over all classes

Expected Benchmark: 40-50%

(Baseline)

Precision & Recall Precision = What percentage of the

positive predictions are correct.
Recall = What percentage of the total
object present were correctly
detected.

Expected Benchmark: > 70% for best

models

Training Loss Lower loss = better training.

If loss plateaus too high, the model
may be underfitting.
If the loss starts increasing, the model
may be overfitting

Expected Benchmark: Should steadily

decrease

Inference Speed Time taken to predict per image

Expected Benchmark: < 50ms per

image

We appreciate your hard work and look forward to reviewing your project!
Feel free to share your work on LinkedIn and show the world your real-world
problem solving skills in action!

15
Tag DualityAI to:
●​ Celebrate your team’s efforts and creativity.
●​ Get noticed by industry experts and recruiters

— The Duality AI Team

16