Hackathon 2024 - Submission of Group // whoami //

Team Members:

• Lou Rainier Gaano
• Saba Nagervadze
• Zhaneta Gasparyan
• Maria Grigoryan
• Mykyta Balandin
• Selin Ilayda Özdemir

---

Project Description

This project focuses on solving a robotic manipulation task using convolutional neural networks (CNNs) integrated with attention mechanisms and U-Net architecture. The model is designed to make precise alignment predictions for manufacturing automation.

We have optimized the model for both accuracy and runtime efficiency to ensure practical applicability in real-world automation scenarios.

---

How to Run

Option 1: Direct Execution

1. Update the paths in the following files to match your machined parts, grippers, and masks:
   • evaluate/task.csv
   • evaluate/ground_truth.csv
2. Run the evaluation script:

   python3 evaluate/eval.py

   or

   python evaluate/eval.py

Option 2: Alternative Script

Run:

python solution/main.py evaluate/task.csv evaluate/ground_truth.csv

Option 3: Docker (Without Visualization)

1. Build the Docker image:

   docker build . -t my-group

2. Run the container:

   docker run -it my-group python evaluate/eval.py

---

Design Decisions

• Architecture:
  We adopted a multi-stage architecture combining CNN layers, attention mechanisms, and U-Net-style upsampling for high-precision predictions. This design allows for more accurate feature extraction and alignment adjustments.

---

Challenges Faced

1. Device Compatibility:
   We encountered issues with MPS devices and dataset variability during training.

2. Resource Limitations:
   Fine-tuning model performance and optimizing runtime were crucial due to resource constraints.

3. Multiprocessing Issues:
   Troubleshooting PyTorch DataLoader multiprocessing (e.g., adjusting the number of workers) was a key challenge. We overcame this by modifying global environment variables and settings for shared memory.

4. Custom Loss Functions:
   We explored custom loss functions to fine-tune the model's response to outliers and out-of-bound predictions, improving overall stability.

---

Features

• Gripper Overlay:
  Adjustable parameters for X, Y, and rotation angle.

• Dataset Creation:
  Tools for generating datasets tailored for machine learning.

---

Additional Notes

• This project demonstrates a practical solution for manufacturing automation, emphasizing both precision and efficiency.