MasterRulesFromChaos

This repository provides code for Master Rules from Chaos: Learning to Reason, Plan, and Interact from Chaos for Tangram Assembly.

Setup

1. Install the required Python packages (PyTorch, torchvision, stable-baselines3, pybullet, numpy and matplotlib). A virtual environment is recommended.
2. Some scripts depend on additional hardware specific packages (e.g. pyrealsense2 for the L515 camera). These are only needed when running on the real robot.

Dataset Generation

To speed up, random patterns for the tangram are pre-generated using the code in the folder util_code

Training

policy_train_rl.py launches training with stable-baselines3. A typical invocation is:

python policy_train_rl.py --algorithm PPO --jobs 4 --n_steps 128 --batch_size 64

The script accepts many arguments for parallel environments, learning rate, number of epochs and more.

Testing

• Reinforcement Learning:
  • Simulation with random layouts: python test_in_sim_random_pattern.py
  • Human-designed layouts: python test_in_sim_human_pattern_auto.py
  • Real robot with L515 depth camera: python test_in_real_with_l515.py (requires hardware and pyrealsense2).

Pretrained RL policies (4-1.zip, 5-1.zip, etc.) can be loaded using stable_baselines3.PPO.load() as demonstrated in the test scripts.

License

This project is licensed under the Apache 2.0 License. See LICENSE for details.