On-Device Reinforcement Learning for Microrobot Control
Analysis
This paper addresses the challenge of controlling microrobots with reinforcement learning under significant computational constraints. It focuses on deploying a trained policy on a resource-limited system-on-chip (SoC), exploring quantization techniques and gait scheduling to optimize performance within power and compute budgets. The use of domain randomization for robustness and the practical deployment on a real-world robot are key contributions.
Key Takeaways
- •Applies reinforcement learning to control a sub-centimeter quadrupedal microrobot.
- •Deploys the RL controller on a resource-constrained SoC (ARM Cortex-M0).
- •Utilizes domain randomization to improve robustness.
- •Investigates integer quantization (Int8) for faster inference.
- •Proposes a resource-aware gait scheduling approach based on power budgets.
Reference
“The paper explores integer (Int8) quantization and a resource-aware gait scheduling viewpoint to maximize RL reward under power constraints.”