Awesome

gym-cassie-run

RL environment (with OpenAI Gym interface) in which a mujoco simulation of Agility Robotics' Cassie robot is rewarded for walking/running forward as fast as possible.

The Cassie model is the one available in DeepMind's mujoco_menagerie, and the gym environment is inspired from HalfCheetah-v4. In addition to the basic reward for forward motion, hand-designed reward signals encourage walking rather than jumping, keeping feet horizontal and not too far apart, and avoiding lateral movements.

The following episode has been obtained after training a TQC agent for 17M timesteps, using the xpag RL platform (at real speed, 100 timesteps = 1 second, cf. top left corner of the gif):

The episode can also be viewed with the view_episode.ipynb notebook in the demo_trajectory/ folder. The trajectory is saved in the demo_trajectory/episode/ folder.

Installation

pip install git+https://github.com/perrin-isir/gym-cassie-run

git clone https://github.com/perrin-isir/gym-cassie-run.git
cd gym-cassie-run

Choose a conda environmnent name, for instance cassierunenv.
The following command creates the cassierunenv environment with the requirements listed in environment.yaml:

conda env create --name cassierunenv --file environment.yaml

If you prefer to update an existing environment (existing_env):

conda env update --name existing_env --file environment.yml

To activate the cassierunenv environment:

conda activate cassierunenv

Finally, to install the gym-cassie-run library in the activated virtual environment:

pip install -e .

Once the installation is complete, you can import the environment in python with:
import gym_cassie_run
This directly registers the environment CassieRun-v0 in gym.