Graph networks for learnable physical simulation

This repository is a partial implementation of Graph networks as learnable physics engines for inference and control.

Dependencies

• DeepMind control suite
• Mujoco
• networkx
• pytorch 0.4.1 (other versions untested)

Generate data

Generate data with gen_data.py script, you should get control signals and resulting 6-link swimmers states.

Train GN

Learn data distribution first with python test_normalizer.py. It will generate normalize.pth. Then run python train_gn.py to train the model. The learning rate schedule corresponds to "fast training" in original paper.

Evaluate GN

python evaluate_gn.py <model path>