/humanoid-gym

Humanoid-Gym: Reinforcement Learning for Humanoid Robot with Zero-Shot Sim2Real Transfer

Primary LanguagePython

Humanoid-Gym: Reinforcement Learning for Humanoid Robot with Zero-Shot Sim2Real Transfer

[Project Page]

Demo

Welcome to our Humanoid-Gym!

Humanoid-Gym is an easy-to-use reinforcement learning (RL) framework based on Nvidia Isaac Gym, designed to train locomotion skills for humanoid robots, emphasizing zero-shot transfer from simulation to the real-world environment. Humanoid-Gym also integrates a sim-to-sim framework from Isaac Gym to Mujoco that allows users to verify the trained policies in different physical simulations to ensure the robustness and generalization of the policies.

This codebase is verified by RobotEra's XBot-S (1.2 meter tall humanoid robot) and XBot-L (1.65 meter tall humanoid robot) in real-world environment with zero-shot sim-to-real transfer.

Features

1. Humanoid Robot Training

This repository offers comprehensive guidance and scripts for the training of humanoid robots. Humanoid-Gym features specialized rewards for humanoid robots, simplifying the difficulty of sim-to-real transfer. In this repository, we use RobotEra's XBot-L as a primary example. It can also be used for other robots with minimal adjustments. Our resources cover setup, configuration, and execution. Our goal is to fully prepare the robot for real-world locomotion by providing in-depth training and optimization.

  • Comprehensive Training Guidelines: We offer thorough walkthroughs for each stage of the training process.
  • Step-by-Step Configuration Instructions: Our guidance is clear and succinct, ensuring an efficient setup process.
  • Execution Scripts for Easy Deployment: Utilize our pre-prepared scripts to streamline the training workflow.

2. Sim2Sim Support

We also share our sim2sim pipeline, which allows you to transfer trained policies to highly accurate and carefully designed simulated environments. Once you acquire the robot, you can confidently deploy the RL-trained policies in real-world settings.

Our simulator settings, particularly with Mujoco, are finely tuned to closely mimic real-world scenarios. This careful calibration ensures that the performances in both simulated and real-world environments are closely aligned. This improvement makes our simulations more trustworthy and enhances our confidence in their applicability to real-world scenarios.

3. Denoising World Model Learning (Coming Soon!)

Denoising World Model Learning(DWL) presents an advanced sim-to-real framework that integrates state estimation and system identification. This dual-method approach ensures the robot's learning and adaptation are both practical and effective in real-world contexts.

  • Enhanced Sim-to-real Adaptability: Techniques to optimize the robot's transition from simulated to real environments.
  • Improved State Estimation Capabilities: Advanced tools for precise and reliable state analysis.

Dexterous Hand Manipulation (Coming Soon!)

Installation

  1. Generate a new Python virtual environment with Python 3.8 using conda create -n myenv python=3.8.
  2. For the best performance, we recommend using NVIDIA driver version 525 sudo apt install nvidia-driver-525. The minimal driver version supported is 515. If you're unable to install version 525, ensure that your system has at least version 515 to maintain basic functionality.
  3. Install PyTorch 1.13 with Cuda-11.7:
    • conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.7 -c pytorch -c nvidia
  4. Install numpy-1.23 with conda install numpy=1.23.
  5. Install Isaac Gym:
    • Download and install Isaac Gym Preview 4 from https://developer.nvidia.com/isaac-gym.
    • cd isaacgym/python && pip install -e .
    • Run an example with cd examples && python 1080_balls_of_solitude.py.
    • Consult isaacgym/docs/index.html for troubleshooting.
  6. Install humanoid-gym:
    • Clone this repository.
    • cd humanoid_gym && pip install -e .

Usage Guide

Examples

# Launching PPO Policy Training for 'v1' Across 8192 Environments
# This command initiates the PPO algorithm-based training for the humanoid task.
python scripts/train.py --task=humanoid_ppo --run_name v1 --headless --num_envs 8192

# Evaluating the Trained PPO Policy 'v1'
# This command loads the 'v1' policy for performance assessment in its environment. 
# Additionally, it automatically exports a JIT model, suitable for deployment purposes.
python scripts/play.py --task=humanoid_ppo --run_name v1

# Implementing Simulation-to-Simulation Model Transformation
# This command facilitates a sim-to-sim transformation using exported 'v1' policy.
python scripts/sim2sim.py --load_model /path/to/logs/XBot_ppo/exported/policies/policy_1.pt

# Run our trained policy
python scripts/sim2sim.py --load_model /path/to/logs/XBot_ppo/exported/policies/policy_example.pt

1. Default Tasks

  • humanoid_ppo

    • Purpose: Baseline, PPO policy, Multi-frame low-level control
    • Observation Space: Variable $(47 \times H)$ dimensions, where $H$ is the number of frames
    • $[O_{t-H} ... O_t]$
    • Privileged Information: $73$ dimensions
  • humanoid_dwl (coming soon)

2. PPO Policy

  • Training Command: For training the PPO policy, execute:
    python humanoid/scripts/train.py --task=humanoid_ppo --load_run log_file_path --name run_name
    
  • Running a Trained Policy: To deploy a trained PPO policy, use:
    python humanoid/scripts/play.py --task=humanoid_ppo --load_run log_file_path --name run_name
    
  • By default, the latest model of the last run from the experiment folder is loaded. However, other run iterations/models can be selected by adjusting load_run and checkpoint in the training config.

3. Sim-to-sim

  • Mujoco-based Sim2Sim Deployment: Utilize Mujoco for executing simulation-to-simulation (sim2sim) deployments with the command below:
    python scripts/sim2sim.py --load_model /path/to/export/model.pt
    

4. Parameters

  • CPU and GPU Usage: To run simulations on the CPU, set both --sim_device=cpu and --rl_device=cpu. For GPU operations, specify --sim_device=cuda:{0,1,2...} and --rl_device={0,1,2...} accordingly. Please note that CUDA_VISIBLE_DEVICES is not applicable, and it's essential to match the --sim_device and --rl_device settings.
  • Headless Operation: Include --headless for operations without rendering.
  • Rendering Control: Press 'v' to toggle rendering during training.
  • Policy Location: Trained policies are saved in humanoid/logs/<experiment_name>/<date_time>_<run_name>/model_<iteration>.pt.

5. Command-Line Arguments

For RL training, please refer to humanoid/utils/helpers.py#L161. For the sim-to-sim process, please refer to humanoid/scripts/sim2sim.py#L169.

Code Structure

  1. Every environment hinges on an env file (legged_robot.py) and a configuration file (legged_robot_config.py). The latter houses two classes: LeggedRobotCfg (encompassing all environmental parameters) and LeggedRobotCfgPPO (denoting all training parameters).
  2. Both env and config classes use inheritance.
  3. Non-zero reward scales specified in cfg contribute a function of the corresponding name to the sum-total reward.
  4. Tasks must be registered with task_registry.register(name, EnvClass, EnvConfig, TrainConfig). Registration may occur within envs/__init__.py, or outside of this repository.

Add a new environment

The base environment legged_robot constructs a rough terrain locomotion task. The corresponding configuration does not specify a robot asset (URDF/ MJCF) and no reward scales.

  1. If you need to add a new environment, create a new folder in the envs/ directory with a configuration file named <your_env>_config.py. The new configuration should inherit from existing environment configurations.
  2. If proposing a new robot:
    • Insert the corresponding assets in the resources/ folder.
    • In the cfg file, set the path to the asset, define body names, default_joint_positions, and PD gains. Specify the desired train_cfg and the environment's name (python class).
    • In the train_cfg, set the experiment_name and run_name.
  3. If needed, create your environment in <your_env>.py. Inherit from existing environments, override desired functions and/or add your reward functions.
  4. Register your environment in humanoid/envs/__init__.py.
  5. Modify or tune other parameters in your cfg or cfg_train as per requirements. To remove the reward, set its scale to zero. Avoid modifying the parameters of other environments!

Troubleshooting

Observe the following cases:

# error
ImportError: libpython3.8.so.1.0: cannot open shared object file: No such file or directory

# solution
# set the correct path
export LD_LIBRARY_PATH="~/miniconda3/envs/your_env/lib:$LD_LIBRARY_PATH" 

# OR
sudo apt install libpython3.8

# error
AttributeError: module 'distutils' has no attribute 'version'

# solution
# install pytorch 1.12.0
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch

# error, results from libstdc++ version distributed with conda differing from the one used on your system to build Isaac Gym
ImportError: /home/roboterax/anaconda3/bin/../lib/libstdc++.so.6: version `GLIBCXX_3.4.20` not found (required by /home/roboterax/carbgym/python/isaacgym/_bindings/linux64/gym_36.so)

# solution
mkdir ${YOUR_CONDA_ENV}/lib/_unused
mv ${YOUR_CONDA_ENV}/lib/libstdc++* ${YOUR_CONDA_ENV}/lib/_unused

Citation

Please cite the following if you use this code or parts of it:

@software{RobotEra2024Humanoid-Gym,
  author = {RobotEra},
  title = {{Humanoid-Gym: Reinforcement Learning for Humanoid Robot with Zero-Shot Sim2Real Transfer}},
  url = {https://github.com/roboterax/humanoid-gym},
  year = {2024}
}

Acknowledgment

The implementation of Humanoid-Gym relies on resources from legged_gym and rsl_rl projects, created by the Robotic Systems Lab. We specifically utilize the LeggedRobot implementation from their research to enhance our codebase.

Any Questions?

If you have further questions, please feel free to contact support@robotera.com.