Project · KaiRA Journal vol.8 · 2024-11
Sim2Real RL for Line-Following Robot
Trained a vision-based RL policy (DrQ-v2) in simulation and achieved zero-shot transfer to a real Raspberry Pi-based robot.
Overview This project tackled the Sim2Real challenge for a line following robot. The goal was to train a policy in simulation using the DrQ v2 algorithm and transfer it, zero shot, to a real world robot. Key Innovations Diverse Simulation: Wrote a program using a Breadth First Search (BFS) algorithm to automatically generate thousands of complex and diverse courses, boosting model robustness. Model Efficiency: Modified the DrQ v2 architecture to accept both image and low dimensional inputs (orientation, velocity), drastically reducing model size for efficient execution on a Raspberry Pi 5. YouTube Demo Also featured in KaiRA Journal vol.8: https://kyoto kaira.github.io/works/collection of journals.html
#Reinforcement Learning #Sim2Real #DrQ-v2 #Robotics #Raspberry Pi