YaojieBao / variant_impedance_tracking_RL

Code for "Learning Quadrupedal Locomotion on Uneven Terrain"

Prerequisite

For building C++

• Mujoco 2.3.5 installed at ~/.mujoco/mujoco235/. (You may change line 26 and 27 of CMakeLists.txt according to your installation of Mujoco).
• glfw3
• X11
• Eigen

For Python environment

• python=3.7
• numpy=1.22.4
• pytorch=1.8.1
• cudatoolkit=11.1
• scipy>=1.5.0
• tensorboard>=2.2.1
• gym
• omegaconf
• hydra-core>=1.1
• rl_games=1.4.0

Build

mkdir build && cd build
cmake ..
make

Run

• To test the environment, enter parallel_sim directory under build and run ./test_a1_env. It will create one simulated environment and run with a trained policy model (which was converted from trained .pt model).
• To train the model, enter parallel_sim directory under build and run python3 train.py task=MujocoA1 headless=True. You may specify max_iteration=XXX to set maximum training epoches (default to 1500).
• To test the trained model under python environment, enter parallel_sim directory under build and run python3 train.py task=MujocoA1 headless=False test=True.

File description

variable_impedance_RL
├── cmake/
├── parallel_sim/
├── converted_models/
├── python/
│   ├── cfg/
│   └── unitree_a1/
├── CMakeLists.txt
└── Readme.md

• parallel_sim/: Contains C++ source files for implementation of parallel simulation environment and python interface.
• converted_models: Contains two binary model file (converted from .pt model, can be read by parallel_sim/eigen_model_from_file.hpp).
• python/: Contains a python wrapper for the parallel simulation environment(mujoco_a1.py).
• python/cfg: Contains configuration of the environment and hyperparameters for RL algorithm.
• python/unitrell_a1/: Contains mujoco model file for Unitree A1 quadrupedal Robot and terrain. The mujoco model for Unitree A1 quadrupedal Robot is copied from mujoco_menagerie.