Solving Complex Dexterous Manipulation Tasks with Trajectory Optimisation and Reinforcement Learning

This repository contains code for our ICML 2021 paper: Solving Complex Dexterous Manipulation Tasks with Trajectory Optimisation and Reinforcement Learning (link to arXiv version). Videos showcasing the obtained results can be found on the main project page. Requirements:

• Mujoco-py
• Pytorch
• Numpy
• joblib
• Dexterous-Gym

Install with pip install -e .

TOPDM contains the code for the trajectory optimisation algorithm. See SCDM/TOPDM/example_experiments.sh for examples of how to run this. Note that this cleaned version of the code seems to be running more slowly than an earlier version - currently looking into this.

TD3_plus_demos contains the code for combining demonstrations with reinforcement learning for the PenSpin task. See SCDM/TD3_plus_demos/run_experiment.sh to run.

We also provide prerun trajectories for all of the environments in SCDM/TOPDM/prerun_trajectories, as well as a file to render these (SCDM/TOPDM/prerun_trajectories/render_demonstrations.py)

We later on also added a version of TOPDM applied to the Humanoid-v3 environment in OpenAI's gym. This is contained in SCDM/TOPDM/humanoid_experiments.