This repository implements a PPO RL agent for a simulated UR5e robotic arm with a gripper to perform pick-and-place tasks. The simulation leverages MuJoCo for physics and rendering, with training designed to process image and proprioceptive inputs.
- Custom Pick-and-Place Environment:
- Designed for UR5e with an RG2 gripper.
- Combines visual and proprioceptive inputs for decision-making.
- Stable-Baselines3 Integration:
- Uses
PPOfrom Stable-Baselines3.
- Uses
- Callbacks:
- Evaluation, checkpoint saving, and stopping based on reward threshold.
- Observation Spaces:
- RGB images (64x64x3).
- Proprioceptive state: joint positions, velocities, gripper state, and TCP position.
This project was inspired and partially adapted from joonhyung-lee/mujoco-robotics-usage.
- Add domain randomization for robustness.
- Use SAC with labeled pick and place tasks