gym-cassie

An OpenAI Gym reinforcement learning interface for Agility Robotics' biped robot Cassie.

Installation

1. Install OpenAI Gym
2. Place mjpro150/ and mjkey.txt in the cassie/cassiemujoco folder (see http://www.mujoco.org/ for info on downloading MuJoCo and obtaining a license).
3. Run $ pip install -e . in root folder (the one that contains setup.py).

Note: I'm currently including a binary of the Cassie MuJoCo C library libcassiemujoco.so in gym_cassie/cassiemujoco (see: https://github.com/osudrl/cassie-mujoco-sim). If you'd like to modify/build that binary from source, note that this repo currently uses a modified version of cassie-mujoco-sim with a foot position API. For now see my fork of the former here: https://github.com/p-morais/cassie-mujoco-sim.

To test installation:

$ python3
> import gym
> import gym_cassie
> env = gym.make("Cassie-v0")
> env.render()
> env.close()

Alternatively:

$ python3
> from gym_cassie import CassieMimicEnv
> env = CassieMimicEnv("walking")
> env.render()

Environments

• Cassie-v0: A "go forward" environment meant to be a close match to OpenAI gym's Humanoid environment in terms of state and reward. Important: This environment is a work in progress and its implementation details will change over time as reward coefficients etc. are empirically determined. Do not yet necessarily expect it to produce a reasonable policy.

• Cassie-mimic-v0: A faithful reimplementation of the environment described in Feedback Control For Cassie With Deep Reinforcement Learning. Also see DeepMimic.

TODO

• Add the option to use the Cassie state estimator output as the observation (this would allow policies to be run on the actual robot).
• Package cassiemujoco as an independent sub-package.