traning_module

Learning

Cleaned up learning codebase for quadruped robot learning

Installation

Then, install usc_learning as a package:

python3 setup.py install --user or pip3 install -e .

Code structure

• [learning] training and testing scripts *learning/rllib_helpers stores policy setup *learning/train_mpc.py train the policy, switch between different envs in here *learning/test_mlp.py test the learnt mlp policy and saves the policy into a .json file for hardware experiment

• [envs] robot class and task environment for trainning *envs/A1.py the robot interface with pybullet (a simplified version of robot class from motion imitation library (https://github.com/erwincoumans/motion_imitation))

  *envs/MPC_* training envrionment of the RL policies to augment MPC for foot reaction, uncertaintiy adaptiaton and agile locomotion. The MPCLocomotion object is called in these envrionment files to update GRF. In the step function, dynamics compensation and swing trajecotry offset are added to the MPC controller.

• [MPC_implementation] the implementation of MPC controller for quadruped robot *mpc_implementation/mpc_osqp.cc A modified solver from motion imitation library (https://github.com/erwincoumans/motion_imitation) that tailor to the proposed fromulation *mpc_implementation/MPCLocomotion.py runs the MPC optimization at desired frequency while updating the swing trajectory for swing foot. Everything is transformed into world/odometry frame to feed to the solver similar to Cheetah-Software (https://github.com/dbdxnuliba/mit-biomimetics_Cheetah).

Train a policy

-cd learning -python3 train_mpc.py

Test the policy

-cd learning -python3 test_mlp.py