hybrid_gait

This repo contain a RL training env for the MIT mini cheetah quadruped robot.

The training task is to train the different gait type according to the various vel input. To perform a walking gait in very low velocity, trot in mid velocity, gallop in high velocity, just like real animal gait trans.

The algorithm stucture is hybrid, which means we use RL to train the 4 gait params(gait_type, gait period, gait offset, gait duration). Then the gait params are as the input to the MPC controller, which calculate the joint force/torque.

We use openai-gym as the RL framework, stable-baselines(PPO) as the RL algorithm lib.

The MPC controller we use the repo of https://github.com/Derek-TH-Wang/quadruped_ctrl

https://github.com/Derek-TH-Wang/hybrid_gait/blob/main/demo.mp4

installation

python

sudo apt-get install virtualenv
virtualenv venv --python=/usr/bin/python3.7
source ${path}/bin/active

pip3 install -i https://pypi.tuna.tsinghua.edu.cn/simple -r requirements.txt

cpp

cd hybrid_gait/quadruped_ctrl/
mkdir build
cd build
cmake ..
make -j

config

see the config/training_params.yaml

mode: train or test
model_file: empty or specific file

run

python3 hybrid_gait/test.py

or use mpi:

mpiexec -n 4 python3 hybrid_gait/test.py