reinforcement-learning

Contents

Shadow-Gym - template used to create a Gym environment for the Shadow Hand.

urdfs - files describing the shape, inertia and mass of various components of the hand

Shadow-Gym/shadow_gym/envs/shadow_env.py - main gymnasium model that interacts with Pybullet. You will mainly modify this file alongside cube.py and hand.py.

logs - each folder in logs contains the tensorboard logs for each run.

models - each folder containes a trained model named according to number of timesteps used e.g. 30000.zip was trained for 30000 timesteps.

train.py - script used to train a model. Automatically generates logs to logs and saves model to models

visualise_model.py - script used to open Pybullet window and view the AI model in action

evaluate_model.py - runs a model 100 times and counts number of times it successfully rotates the cube to the target orientation

Shadow-Gym/shadow-gym/resources/cube_colour_visualiser.py - utility script to view the hand environment with gravity turned off

Installation Instructions

(Recommended) Create a virtual environment with conda or venv

Enter the Shadow-Gym folder (folder with setup.py)
Run pip install -e . to install other required packages for the hand model

For the AI modules run pip install stable_baselines3 tensorboard

Github recommendation

Try following this training loop

1. Modify environment and training code
2. Train model
3. Visualise and evalute model
4. Commit all files (model, logs and environment code) to git
5. ONLY AFTER COMMITTING, modify environment/training code