Humanoidbench RL#

We provide a basic RL training example for Humanoidbench tasks.

RL framework: stable-baselines3

RL algorithm: PPO

Simulator: MuJoCo and IsaacGym

Installation#

pip install stable-baselines3
pip install wandb
pip install wandb[media]
pip install tensorboard

Wandb login, enter your wandb account token.

wandb login

Training#

python roboverse_learn/humanoidbench_rl/train_sb3.py --sim mujoco --num_envs 1 --robot=h1 --task humanoidbench:Stand --wandb_entity <your_wandb_entity_name> --use_wandb

python roboverse_learn/humanoidbench_rl/train_sb3.py --sim isaacgym --num_envs 2 --robot=h1 --task humanoidbench:Stand --wandb_entity <your_wandb_entity_name> --use_wandb

python roboverse_learn/humanoidbench_rl/train_sb3.py --sim isaaclab --num_envs 2 --robot=h1 --task humanoidbench:Stand --wandb_entity <your_wandb_entity_name> --use_wandb

Warning#

  • Isaacgym has problem now, see issue:

    • https://github.com/RoboVerseOrg/RoboVerse/issues/61, corresponding to the code: metasim/cfg/robots/h1_cfg.py

      • Switch the fix_base_link to True to ignore this bug.

    • https://github.com/RoboVerseOrg/RoboVerse/issues/57, corresponding to the code: metasim/utils/humanoid_robot_util.py

      • Switch the head_body to mid360_link to ignore this bug.

  • Isaaclab has problem now:

    • H1 usd doesn’t have head body, see code: metasim/utils/humanoid_robot_util.py

      • Switch the head_body to mid360_link to ignore this bug.

Current implementation of Isaacgym and Isaaclab has problem, after dealing with these 2 bugs or ignoring them, it’s running well.

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