Dual-Arm Space Robotic Manipulation with Deep Reinforcement Learning
A PyBullet-based dual-arm robotic manipulation framework using PPO and SAC for coordinated target reaching, collision avoidance, and noisy-observation robustness, with task-specific reward functions and multi-objective evaluation metrics.
Developed a PyBullet-based dual-arm robotic manipulation framework using PPO and SAC for coordinated target reaching, collision avoidance, and noisy-observation robustness. The system is designed around the challenges of space robotics: limited sensor accuracy, no ground truth pose, and the need for collision-free coordinated motion between two arms.
Overview
Space robotic manipulators face unique challenges — microgravity dynamics, sensor noise from harsh environments, and the inability to rely on precise feedback. This project addresses these constraints using deep RL agents that learn robust manipulation policies directly from noisy simulated observations.
Key Contributions
- Dual-Arm Coordination: Designed joint policies for two robot arms operating on a shared workspace, trained to reach targets while avoiding self-collision and inter-arm interference
- Noisy Observations: Injected observation noise during training to improve sim-to-real robustness in sensor-degraded conditions
- Algorithm Comparison: Benchmarked PPO (on-policy) vs. SAC (off-policy, sample efficient) on the same task suite
- Reward Design: Formulated task-specific rewards combining target proximity, smoothness penalties, collision penalties, and success bonuses
- Evaluation Metrics: Designed multi-objective metrics covering success rate, path efficiency, joint effort, and robustness to observation perturbation
- Randomized Test Scenarios: Evaluated on randomized initial conditions and target placements to test generalization