MimicKit for Isaac Lab

Physics-based character animation using Isaac Lab, featuring motion imitation with DeepMimic, AMP, and ASE algorithms.

This is an Isaac Lab implementation of MimicKit, adapted for NVIDIA Isaac Sim and Isaac Lab. The G1 robot configuration and USD assets are adapted from TrackerLab.

Note: This repository was developed with significant AI assistance. If you find areas for improvement or optimization, please open an issue or pull request. Contributions are welcome!

Features

• DeepMimic: Learn locomotion skills from reference motion data
• AMP: Style-aware motion learning with discriminator
• ASE: Diverse skill discovery through exploration
• SMPL & G1 Support: Humanoid (SMPL) and robot (Unitree G1) character models
• GPU-Accelerated: Parallel simulation with NVIDIA Isaac Sim

IsaacLab Demos (Simulation)

I reconstruct the scene using 3D Gaussian Splatting combined with 3DGRUT

Installation

Prerequisites

• Ubuntu 20.04/22.04/24.04
• NVIDIA GPU with CUDA support
• Isaac Lab 2.1+
• Python 3.10+
• Isaac Sim 4.5+ (tested with 5.0)

Setup

1. Install Isaac Lab following official instructions

2. Clone and install MimicKit for Isaac Lab:

git clone https://github.com/nathanwu7/MimicKit_IsaacLab.git
cd MimicKit_IsaacLab
pip install -e source/mimickit_isaaclab

3. Download motion data:

# Motion data should be placed in:
# - data/motions/smpl/*.pkl  (SMPL motions)
# - data/motions/g1/*.pkl    (G1 motions)
# Character models in:
# - data/assets/smpl/        (SMPL models)
# - data/assets/g1/          (G1 models)

Quick Start

Training

DeepMimic

# SMPL humanoid
python scripts/train.py --task MimicKit-DeepMimic-SMPL-Direct-v0

# G1 robot
python scripts/train.py --task MimicKit-DeepMimic-G1-Direct-v0

AMP (Adversarial Motion Priors)

# SMPL humanoid
python scripts/train.py --task MimicKit-AMP-SMPL-Direct-v0

# G1 robot
python scripts/train.py --task MimicKit-AMP-G1-Direct-v0

ASE (Adversarial Skill Embeddings)

# SMPL humanoid
python scripts/train.py --task MimicKit-ASE-SMPL-Direct-v0

# G1 robot
python scripts/train.py --task MimicKit-ASE-G1-Direct-v0

Training Options

python scripts/train.py --task MimicKit-DeepMimic-G1-Direct-v0 \
    --num_envs 4096 \          # Number of parallel environments
    --max_iterations 10000 \   # Training iterations
    --headless                 # Run without GUI (faster)

Inference & Visualization

1. Play Trained Policy (Strategy Only)

python scripts/play.py --task MimicKit-DeepMimic-G1-Direct-v0 \
    --num_envs 1 \
    --checkpoint path/to/model.pt

2. Replay Reference Motion (Data Only)

python scripts/replay.py --task MimicKit-DeepMimic-G1-Direct-v0 \
    --num_envs 1 \
    --motion_file data/motions/g1/g1_walk.pkl

3. Compare Policy with Reference (Side-by-Side)

python scripts/compare.py --task MimicKit-DeepMimic-G1-Direct-v0 \
    --num_envs 1 \
    --checkpoint path/to/model.pt

Available Environments

SMPL Humanoid (69 DOF)

Environment ID	Description
MimicKit-DeepMimic-SMPL-Direct-v0	DeepMimic with SMPL humanoid
MimicKit-AMP-SMPL-Direct-v0	AMP with SMPL humanoid
MimicKit-ASE-SMPL-Direct-v0	ASE with SMPL humanoid

Unitree G1 Robot (29 DOF)

Environment ID	Description
MimicKit-DeepMimic-G1-Direct-v0	DeepMimic with G1 robot
MimicKit-AMP-G1-Direct-v0	AMP with G1 robot
MimicKit-ASE-G1-Direct-v0	ASE with G1 robot

Configuration

Environment Configuration

Located in source/mimickit_isaaclab/tasks/direct/{algorithm}/{algorithm}_{robot}_env_cfg.py

Key Parameters:

# Motion file
motion_file = "data/motions/g1/g1_walk.pkl"

# Termination conditions
enable_early_termination = True
termination_height = 0.1         # Min height before termination
pose_termination = True
pose_termination_dist = 5.0      # Max deviation from reference

# Visualization
visualize_ref_char = False       # Show reference character (training)
ref_char_offset = [2.0, 0.0, 0.0]  # Reference character offset

Agent Configuration

Located in source/mimickit_isaaclab/tasks/direct/{algorithm}/agents/

Motion Data Format

Motion files (.pkl) should contain a dictionary with:

{
    "fps": 30,                    # Frames per second
    "loop_mode": "wrap",          # "wrap" or "clamp"
    "frames": {
        "root_pos": np.array,     # Shape: (num_frames, 3)
        "root_rot_exp_map": np.array,  # Shape: (num_frames, 3)
        "joint_dof": np.array,    # Shape: (num_frames, num_dofs)
    }
}

DOF Configuration

• SMPL: 69 DOF (3 DOF per joint, spherical joints)
• G1: 29 DOF (1 DOF per joint, revolute joints)

Reference

@misc{mimickit_peng2025,
  title={MimicKit: A Reinforcement Learning Framework for Motion Imitation and Control}, 
  author={Xue Bin Peng},
  year={2025},
  eprint={2510.13794},
  archivePrefix={arXiv},
  primaryClass={cs.GR},
  url={https://arxiv.org/abs/2510.13794}, 
}

@article{2018-TOG-deepMimic,
  author = {Peng, Xue Bin and Abbeel, Pieter and Levine, Sergey and van de Panne, Michiel},
  title = {DeepMimic: Example-guided Deep Reinforcement Learning of Physics-based Character Skills},
  journal = {ACM Trans. Graph.},
  volume = {37},
  number = {4},
  year = {2018},
  doi = {10.1145/3197517.3201311},
}

@article{2021-TOG-AMP,
  author = {Peng, Xue Bin and Ma, Ze and Abbeel, Pieter and Levine, Sergey and Kanazawa, Angjoo},
  title = {AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control},
  journal = {ACM Trans. Graph.},
  volume = {40},
  number = {4},
  year = {2021},
  doi = {10.1145/3450626.3459670},
}

@article{2022-TOG-ASE,
  author = {Peng, Xue Bin and Guo, Yunrong and Halper, Lina and Levine, Sergey and Fidler, Sanja},
  title = {ASE: Large-scale Reusable Adversarial Skill Embeddings for Physically Simulated Characters},
  journal = {ACM Trans. Graph.},
  volume = {41},
  number = {4},
  year = {2022},
}
@software{zheng2025@trackerLab,
  author = {Ziang Zheng},
  title = {TrackerLab: One step unify IsaacLab with multi-mode whole-body control.},
  url = {https://github.com/interval-package/trackerLab},
  year = {2025}
}

Citation

@software{MimicKit_IsaacLab,
  author = {Qiwei Wu},
  title = {MimicKit Support for Isaac Lab},
  url = {https://github.com/NathanWu7/MimicKit_IsaacLab},
  year = {2025}
}

Or star my repo.

License

This project is licensed under the BSD-3-Clause License. See LICENSE file for details.

Acknowledgments

• Based on MimicKit by Xue Bin Peng
• Built on Isaac Lab by NVIDIA
• G1 robot configuration and USD assets adapted from TrackerLab by Ziang Zheng and this repo https://github.com/Beat-in-our-hearts/mujoco_sim2sim_assets
• Inspired by:
  • DeepMimic (Peng et al., 2018)
  • AMP (Peng et al., 2021)
  • ASE (Peng et al., 2022)

Contact & Support

For issues, questions, or contributions:

• Open an issue on GitHub
• Refer to Isaac Lab documentation for simulation-specific questions
• See original MimicKit for Isaac Gym implementation