In this repo, we offer two baseline methods for simple humanoid-object interaction task. Feel free to test our designed HOI task Success Rate (SR%) in simulator for your research. If you want to apply our framework to your task setting, we also release our training code:

1. RSS 2026: HAIC: Humanoid Agile Object Interaction Control via Dynamics-Aware World Model [project page] [arXiv] [code]

2. Arxiv 2026: VAIC: Vision-Guided Humanoid Agile Object Interaction Control via Decoupled Commands [project page] [arXiv] [code]

Skateboard

Box via terrains

• Release the HAIC policy for its sim2sim performance on HOI task.
• Release the VAIC policy for its sim2sim performance on HOI task.

Use the project Python environment, then install the runtime packages:

conda create -n deploy_onnx python=3.8
conda activate deploy_onnx
pip install mujoco onnxruntime tqdm imageio imageio-ffmpeg opencv-python

Run a specific task:

python scripts/deploy_haic.py --task box_plane
python scripts/deploy_haic.py --task box_slopestair
python scripts/deploy_haic.py --task box_stairslope
python scripts/deploy_haic.py --task pull_cart
python scripts/deploy_haic.py --task pull_cart_box
python scripts/deploy_haic.py --task push_cart
python scripts/deploy_haic.py --task push_wheelchair
python scripts/deploy_haic.py --task skateboard

python scripts/deploy_haic.py \
  --task skateboard \
  --sim_dt 0.002 \
  --control_dt 0.02 \
  --viewer_dt 0.01 \
  --num_actuators 29

• --sim_dt: MuJoCo physics timestep.
• --control_dt: policy/control timestep. Must be an integer multiple of --sim_dt.
• --viewer_dt: viewer/video render interval.
• --num_actuators: expected robot actuator count.
• --record_video: save a 2x-speed MP4 named <task>_sim2sim.mp4.
• --video_width: recorded video width. Default: 1280.
• --video_height: recorded video height. Default: 720.
• --video_codec: recorded video codec. Default: h264.
• --no_real_time: run as fast as possible.
• --check_assets: print task triplets and exit after validation.

Record video:

python scripts/deploy_haic.py --task skateboard --record_video

Fast H.264 recording helper:

python scripts/record_haic.py --task skateboard

Disable real-time sleeping:

python scripts/deploy_haic.py --task skateboard --no_real_time

Check that every configured task has its policy, motion, and XML assets:

python scripts/deploy_haic.py --check_assets

Deploy_HAIC/
├── scripts/
│   ├── deploy_haic.py          ← deploy HAIC policy in MuJoCo
│   └── depoly_vaic.py          ← deploy VAIC policy in MuJoCo
├── controller/
│   ├── policy/
│   │   ├── haic/               ← HAIC ONNX policies 
│   │   └── vaic/               ← VAIC ONNX policies
│   └── motion/
│       └── mirobotA/           ← reference motion folders with meta.json and motion.npz
├── assets/
│   └── robots/
│       └── g1/
│           ├── meshes_v1/      ← G1 mesh assets used by the XML files
│           └── *.xml           ← MuJoCo task XML files
└── readme.md

scripts/deploy_haic.py loads one task triplet:

• ONNX policy from controller/policy/haic/
• reference motion from controller/motion/mirobotA/
• MuJoCo XML from assets/robots/g1/

The G1 XML files use mesh assets from assets/robots/g1/meshes_v1/.

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}
}

@article{li2026vaic,
  title = {VAIC: Vision-Guided Humanoid Agile Object Interaction Control via Decoupled Commands},
  author = {Li, Dongting and Wu, Qianyang and Chen, Xingyu and Li, Liang and Lin, Yuhang and Wu, Sikai and Zhang, Guoyao and Zhou, Mingliang and Xiang, Diyun and Zhang, Qiang and Xu, Renjing and Ma, Jianzhu},
  journal = {arXiv preprint arXiv:2606.09286},
  year = {2026}
}