opt-mimic-raisim

October 5, 2022 ยท View on GitHub

Reinforcement Learning (RL) code for the paper "OPT-Mimic: Imitation of Optimized Trajectories for Dynamic Quadruped Behaviors". This repository was modified from the raisimLib repository. Modifications include the addition of a URDF file for the ODRI Solo 8 robot to rsc/solo8_v7/, and the addition of RL environment and training code to raisimGymTorch/raisimGymTorch/env/envs/solo8_env/.

Installation

The installation steps from Raisim should be followed first. Then this repo can be cloned into your $WORKSPACE directory.

Training RL Policies

  • After any changes to C++ code, go to raisimGymTorch/ and run python setup.py develop
  • Go to raisimGymTorch/raisimGymTorch/env/envs/solo8_env/ and run python vec_ppo.py -n my-experiment-name to start RL training
  • Files indicating training progress will be saved to raisimLibSolo/raisimGymTorch/raisimGymTorch/env/envs/solo8_env/stats/my-experiment-name/
  • The reference motion to track can be specified in the ref_filename argument of raisimLibSolo/raisimGymTorch/raisimGymTorch/env/envs/solo8_env/cfg.yaml. this should correspond to a filename in raisimGymTorch/raisimGymTorch/env/envs/solo8_env/traj/, which includes reference motion csv files produced from trajectory optimization
  • Note that Raisim comes bundled with RL training code building off of OpenAI Stable Baselines, but this is unused and instead a custom implementation of RL training code is used here

Testing a trained RL policy

  • Go to raisimGymTorch/raisimGymTorch/env/envs/solo8_env/ and run python test_policy.py my-experiment-name/latest.pt to run the latest policy trained using python vec_ppo.py -n my-experiment-name
  • Early termination, which is important during RL training, can be turned off during testing by setting cfg['environment']['disable_termination'] to true in raisimGymTorch/raisimGymTorch/env/envs/solo8_env/test_policy.py (line 40 as of writing).