HoRD Dataset: Processed Motion Data for Robust Humanoid Control

This dataset card describes the processed training data used by HoRD (History-Conditioned Reinforcement Learning and Online Distillation), a two-stage framework for robust humanoid motion 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

Overview

The HoRD dataset provides processed motion data for humanoid policy training. It is designed for the HoRD teacher-student pipeline:

• Stage 1 (Teacher RL): train an expert policy with privileged observations and domain randomization.
• Stage 2 (Online Distillation): distill to a deployable student policy with sparse commands and partial observations.

In HoRD, this data supports robust motion imitation and zero-shot transfer experiments across simulators.

Dataset Contents

Current primary release:

• train_g1_all.pt: processed motion data used for training and evaluation configuration.

Quick Start

Install Hugging Face CLI:

pip install -U "huggingface_hub[cli]"

Download the dataset file:

mkdir -p data
huggingface-cli download --repo-type=dataset tony0517/HoRD train_g1_all.pt --local-dir data --local-dir-use-symlinks False

Use in HoRD training/evaluation:

motion_file=data/train_g1_all.pt

Example Training Commands

Stage 1:

python hord/train_agent.py +exp=full_body_tracker/transformer +robot=g1 +simulator=isaaclab motion_file=data/train_g1_all.pt +experiment_name=full_body_tracker_g1 ++headless=True

Stage 2:

python hord/train_agent.py +exp=masked_mimic/no_vae +robot=g1 +simulator=isaaclab motion_file=data/train_g1_all.pt +experiment_name=masked_mimic_g1 ++headless=True

License

This dataset card is released under the MIT License.
Please also follow the licenses and terms of any upstream data sources used in preprocessing.

Citation

If you find this dataset 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}
}