Bipedal point-to-point walk training code

Overview

This repository shows the training code for paper "Learning Point-to-Point Bipedal Walking Without Global Navigation" with Isaaclab.

Installation

• Install Isaac Lab by following the installation guide. Please use Isaaclab v2.1.0 with IsaacSim 4.5.0

• Using a python interpreter that has IsaacLab installed, install the library

cd Bipedal-p2p-walk
python -m pip install -e source/AzureLoong

Run

Training and play an agent with RSL-RL on a bipedal robot AzureLoong:

python scripts/rsl_rl/train.py --task=walk_p2p_s1 --headless
python scripts/rsl_rl/play.py --task=walk_p2p_s1 --num_envs 5

Project Structure

The project is mainly organized as follows:

Bipedal-p2p-walk/
├── cmd.txt
├── scripts/
│   └── rsl_rl/
│       ├── cli_args.py
│       ├── export.py
│       ├── play.py
│       └── train.py
└── source/
    └── AzureLoong/
        ├── AzureLoong/
            ├── assets/
            │   ├── AzureLoong.py
            │   ├── __init__.py
            │   └── Robots/
            │       ├── AzureLoong_shortFeet.usd
            │       └── configuration/
            ├── tasks/
                └── flat_walk/
                    ├── agents/
                    │   └── rsl_rl_ppo_cfg.py
                    ├── base_scripts/
                    │   ├── cfg_base.py
                    │   └── env_base.py
                    ├── cfg_p2p_s1.py
                    ├── cfg_p2p_s2.py
                    └── env_p2p_s1.py

The robot asset file is store as AzureLoong_shortFeet.usd.

Joint configurations such as stiffness and damping are configured in AzureLoong.py.

Environment settings, reward functions and corresponding scales are defined in env_p2p_s1.py and cfg_p2p_s1.py. cfg_p2p_s2.py is the second stage training config with more domain randomization.

PPO parameters are configured in rsl_rl_ppo_cfg.py.

References:

Humanoid-Gym: Reinforcement Learning for Humanoid Robot with Zero-Shot Sim2Real Transfer. https://github.com/roboterax/humanoid-gym