Humanoidbench RL

We provide a basic RL training example for Humanoidbench tasks.

RL framework: stable-baselines3

RL algorithm: PPO

Simulator: MuJoCo and IsaacGym and IsaacLab

Installation

pip install stable-baselines3
pip install wandb
pip install tensorboard

Wandb login, enter your wandb account token.

wandb login

Training

NOTE: You can add --use_wandb --wandb_entity <your_wandb_entity_name> to use wandb to log the training process.

• MuJoCo:

  python roboverse_learn/humanoidbench_rl/train_sb3.py --sim mujoco --num_envs 1 --robot=h1 --task humanoidbench:Stand
  

• IsaacGym:

  python roboverse_learn/humanoidbench_rl/train_sb3.py --sim isaacgym --num_envs 2 --robot=h1 --task humanoidbench:Stand
  

• IsaacLab:

  python roboverse_learn/humanoidbench_rl/train_sb3.py --sim isaaclab --num_envs 2 --robot=h1 --task humanoidbench:Stand
  

Task list

• Balance

  • Not Implemented because of collision detection issues.

• Crawl

• Cube

• Door

• Highbar

  • Not implemented due to the need to connect H1 and Highbar.

• Hurdle

  • Not Implemented because of collision detection issues.

• Maze

  • Not Implemented because of collision detection issues.

• Package

• Pole

  • Not Implemented because of collision detection issues.

• Powerlift

• Push

• Run

• Sit

• Slide

• Spoon

  • Not Implemented because of sensor detection issues.

• Stair

• Stand