HoRD Models: Checkpoints for Robust Humanoid Motion Control

This model card describes the released checkpoints for HoRD (History-Conditioned Reinforcement Learning and Online Distillation), a two-stage framework for robust humanoid control under domain shift.

• Paper: HoRD: Robust Humanoid Control via History-Conditioned Reinforcement Learning and Online Distillation
• Project Page: https://tonywang-0517.github.io/hord/
• Code Repository: https://github.com/tonywang-0517/hord
• Dataset Repository: https://huggingface.co/datasets/tony0517/HoRD
• Model Repository: https://huggingface.co/tony0517/HoRD

Model Overview

HoRD checkpoints are trained for robust humanoid motion imitation and control with online adaptation capability.

• Stage 1 (Teacher RL): an expert policy trained with privileged observations and domain randomization.
• Stage 2 (Online Distillation): a deployable student policy distilled from the teacher under partial observability and sparse commands.

These checkpoints are intended for evaluation and downstream experiments in IsaacLab and Genesis settings.

Model Contents

Typical release artifact:

• your_checkpoint.ckpt: pretrained HoRD checkpoint for evaluation or fine-tuning.

Quick Start

Install Hugging Face CLI:

pip install -U "huggingface_hub[cli]"

Download a checkpoint from the model repository:

mkdir -p results
huggingface-cli download tony0517/HoRD your_checkpoint.ckpt --local-dir results --local-dir-use-symlinks False

Use in HoRD evaluation:

+checkpoint=results/your_checkpoint.ckpt

Example Evaluation Command

python hord/eval_agent.py +exp=full_body_tracker/transformer +robot=g1 +simulator=isaaclab +motion_file=data/train_g1_all.pt +experiment_name=full_body_tracker ++headless=False +checkpoint=results/your_checkpoint.ckpt ++num_envs=1

License

This model card is released under the MIT License.

Citation

If you find these checkpoints useful, please cite:

@article{wang2026hord,
  title={HoRD: Robust Humanoid Control via History-Conditioned Reinforcement Learning and Online Distillation},
  author={Wang, Puyue and Hu, Jiawei and Gao, Yan and Wang, Junyan and Zhang, Yu and Dobbie, Gillian and Gu, Tao and Johal, Wafa and Dang, Ting and Jia, Hong},
  journal={arXiv preprint arXiv:2602.04412},
  year={2026}
}