This repository hosts the open-source release for the RSS 2026 paper HAIC: Humanoid Agile Object Interaction Control via Dynamics-Aware World Model.

• [2026-06-02] Release the asset module (active_adaptation/assets): the G1 robot USDs, the interaction object USDs, and the asset configuration code.
• [2026-06-08] Initial public release: environment and task configurations, dynamics-aware world model, training code, and evaluation/play scripts.
• [2026-06-10] Release MuJoCo Sim2Sim deployment code via the dedicated open-source repository HOIC-baseline for real-world policy transfer.

• Release the asset module (G1 robot USDs, object USDs, and asset configuration)
• Release the environment and task configurations
• Release the dynamics-aware world model implementation
• Release the training code
• Release the evaluation and play scripts
• Release the sim-to-sim deployment code
• Provide setup instructions and usage documentation

# setup conda environment
conda create -n haic python=3.11 -y
conda activate haic

# install isaacsim
pip install "isaacsim[all,extscache]==5.1.0" --extra-index-url https://pypi.nvidia.com
isaacsim # test isaacsim

# install isaaclab
cd ..
git clone git@github.com:isaac-sim/IsaacLab.git
cd IsaacLab
git checkout v2.3.2
./isaaclab.sh -i none

# install haic
cd ..
git clone https://github.com/ldt29/HAIC
cd HAIC
pip install -e .

Visualize motions in Isaac Sim with task.command.replay_motion=true:

python scripts/play.py algo=ppo_haic_train task=G1/haic/skateboard task.command.replay_motion=true

Teacher policy

# train policy
python scripts/train.py algo=ppo_haic_train task=G1/haic/skateboard
# evaluate policy
python scripts/play.py algo=ppo_haic_train task=G1/haic/skateboard checkpoint_path=run:<wandb-run-path>

Student policy

# train policy
python scripts/train.py algo=ppo_haic_finetune task=G1/haic/skateboard checkpoint_path=run:<student_wandb-run-path>
# evaluate policy
python scripts/play.py algo=ppo_haic_finetune task=G1/haic/skateboard checkpoint_path=run:<student_wandb-run-path>

To export trained policies, add export_policy=true to the play script.

We provide MuJoCo-based Sim2Sim deployment code to facilitate the transfer of agile humanoid object interaction policies from Isaac Sim to real-world platforms, whch is open-sourced at:

• GitHub: Cybercal/HOIC-baseline

This repository includes:

• MuJoCo Simulation Environment: A complete simulation setup tailored for humanoid-object interaction tasks.
• Deployment & Inference Pipeline: Lightweight scripts supporting direct deployment and evaluation of pre-trained policy ONNX models.
• Sim2Sim Validation Tools: Benchmarks and evaluation utilities to bridge the gap between Isaac Sim and real-world deployment.

The simulation assets live under active_adaptation/assets: the G1 robot USDs, the interaction object USDs, and the Python configuration code that wires them up for simulation. See active_adaptation/assets/README.md for the full asset list and usage details.

This repository is built on top of HDMI: Learning Interactive Humanoid Whole-Body Control from Human Videos. We thank the authors for open-sourcing their work.

If you find our work useful for your research, please consider citing us:

@article{li2026haic,
  title = {HAIC: Humanoid Agile Object Interaction Control via Dynamics-Aware World Model},
  author = {Li, Dongting and Chen, Xingyu and Wu, Qianyang and Chen, Bo and Wu, Sikai and Wu, Hanyu and Zhang, Guoyao and Li, Liang and Zhou, Mingliang and Xiang, Diyun and Ma, Jianzhu and Zhang, Qiang and Xu, Renjing},
  journal = {arXiv preprint arXiv:2602.11758},
  year = {2026}
}