This repository contains the code for the training framework of the RL_WBC project. The control framework consists of a high-level RL policy and a low-level Whole-Body Control (WBC) policy.

Acknowledgement: The training framework is partially based on the rsl_rl and cajun.

python train.py

Training arguments:

• --num_envs=4096: Number of parallel environments.
• --show_gui=False: Show the GUI of the simulation.
• --max_iterations=600: Maximum number of iterations.
• --friction_type=[pyramid,cone]: Friction type of the WBC formulation.
• --solver_type=[pdhg, qpth]: QP solver type.
• --env_dt=0.02: RL policy time step.

python eval.py

Evaluation arguments:

• --num_envs=4096: Number of parallel environments.
• --use_gpu=True: Use GPU for evaluation.
• --show_gui=False: Show the GUI of the simulation.
• --use_real_robot=[0,1,2]: 0 for isaac gym, 1 for mujoco and 2 for real robot.
• --name=[go1,go2,a1,mini_cheetah]: Robot name.

python wbc_example.py