Datasets:

leggedrobotics
/

egohtr

ArXiv:

License:

Dataset card Files Files and versions

xet

Community

You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

The DATASET is made available under the CC-BY-4.0 license, which permits commercial use.
By downloading this dataset, you agree to the following terms:

Attribution: You must appropriately cite the DATASET in any publication, product documentation, or service that utilizes it. ETH Zurich retains the right to disclose and advertise the DATASET use by any licensee.
Privacy & Identity: You shall not carry out any procedures with the DATASET (linking, comparison, processing) with which the identity of a person could be derived. The de-anonymization of the DATASET is strictly prohibited.
Updates & Removals: The DATASET may be modified at any moment to have deleted, changed, or added information. If requested, you agree to update your local copy and delete specific data points.
Liability: The DATASET is provided “as is” without warranty of any kind, express or implied. In no event shall the authors, ETH Zurich, or licensors be liable for any claims, damages, or other liability arising from the use of the DATASET.
Jurisdiction: These terms shall be governed by the laws of Switzerland.

Dataset Card for EgoHTR

Updates

30.10.2026 added retargeted G1 motions, fixed footlock data leak and minor motion sequence mistakes

Dataset Details

Dataset Description

EgoHTR (Egocentric Human-Terrain Reconstruction) is a multimodal dataset of 4D human motion precisely aligned to complex, unstructured terrain. It combines egocentric wearables (Project Aria) and IMU-based motion capture with portable 3D scene scanning to reconstruct human motion in diverse indoor and outdoor environments — enabling terrain-aware humanoid locomotion research where existing datasets, limited to controlled or unaligned capture, fall short.

Curated by: Alex Brandes and co-authors, ETH Zürich
Modality: Egocentric RGB video, Exocentric RGB video, IMU motion capture, 3D scene scans, optimized SMPL-X body models

Dataset Sources

Paper: EgoHTR: Egocentric 4D Demonstrations of Human Terrain Traversal (arXiv:2607.13472), accepted at CoRL 2026
Project page: https://egohtr.github.io
Generation pipeline (code): Github

Direct Use

Benchmarking 4D human–scene reconstruction and human mesh recovery methods, evaluated against MoCap ground truth
Training and evaluating perceptive, terrain-aware whole-body control / humanoid locomotion policies (demonstrated on a Unitree G1)
Retargeting human motion to robot embodiments via the accompanying pipeline

Download

The full repository is large (~700 GB, mostly images inside full_sequence.npz). Use --include to download only what you need:

pip install -U huggingface_hub

# reconstructions of all sequences (full_sequence.npz + full_opt_sequence.npz)
hf download leggedrobotics/egohtr --repo-type dataset --local-dir egohtr \
    --include "sequences/*/full_sequence.npz" "sequences/*/full_opt_sequence.npz"

# only the optimized SMPL-X motions (small, no images)
hf download leggedrobotics/egohtr --repo-type dataset --local-dir egohtr \
    --include "sequences/*/full_opt_sequence.npz"

# a single sequence
hf download leggedrobotics/egohtr --repo-type dataset --local-dir egohtr \
    --include "sequences/parkour_s1_gym_04/full_sequence.npz" "sequences/parkour_s1_gym_04/full_opt_sequence.npz"

# retargeted robot motions
hf download leggedrobotics/egohtr --repo-type dataset --local-dir egohtr --include "retargeted/g1_gmr_isaaclab/*"
hf download leggedrobotics/egohtr --repo-type dataset --local-dir egohtr --include "retargeted/g1_omniretarget_mujoco/*"

Dataset Structure

Each sequence folder contains the compressed processed outputs per sequence: full_sequence.npz (complete synchronized data), full_opt_sequence.npz, (optional: full_retarget_sequence.npz). Full field-level schema is documented in the EgoHTR pipeline repository.

Additionally it contains the raw data of the sensors (temporally and spatially unaligned):

aria/ — egocentric RGB + SLAM trajectory + hand tracking (Project Aria)
egomc/ — IMU-suit motion capture (BVH + raw)
exomc/ — optional marker-based MoCap, used for ground-truth evaluation
scene/ — 3D scanner point cloud and textured mesh
retarget/ — motion retargeted to robot embodiments (URDF/XML) (earlier, partial retargeting; superseded by retargeted/)

Retargeted motions (`retargeted/`)

All 55 sequences retargeted to the 29-DoF Unitree G1 at full length (30 fps), scaled to robot size (1.32 / 1.78), with the scanned scene cropped around the motion (3 m radius):

g1_gmr_isaaclab/: GMR retargeting with USD terrains for Isaac Lab (triangle-mesh collider).
g1_omniretarget_mujoco/: OmniRetarget interaction-mesh retargeting with non-penetration against a convex (CoACD) scene, as MuJoCo XML.


GMR

OmniRetarget + CoACD

Formats and known issues are described in the README of each folder. Thanks to the authors of GMR, OmniRetarget and CoACD for their open-source code.

Citation

Please cite our dataset in case of usage.

BibTeX:

@misc{brandes2026egohtregocentric4ddemonstrations,
      title={EgoHTR: Egocentric 4D Demonstrations of Human Terrain Traversal}, 
      author={Alex Brandes and Haig Conti Georges Sajelian and Manthan Patel and Dominik Hollidt and Chenhao Li and Matthias Heyrman and Oliver Hausdoerfer and Manuel Kaufmann and Xi Wang and Jonas Frey and Angela P. Schoellig and Christian Holz and Marc Pollefeys and Marco Hutter},
      year={2026},
      eprint={2607.13472},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2607.13472}, 
}

APA:

Brandes, A., Sajelian, H. C. G., Patel, M., Hollidt, D., Li, C., Heyrman, M., Hausdoerfer, O., Kaufmann, M., Wang, X., Frey, J., Schoellig, A. P., Holz, C., Pollefeys, M., & Hutter, M. (2026). EgoHTR: Egocentric 4D demonstrations of human terrain traversal. In Conference on Robot Learning (CoRL).

Contact

If you have any questions please open an issue on the Github repo or contact Alex Brandes

Downloads last month: 7,269

Total file size:

763 GB

Paper for leggedrobotics/egohtr

EgoHTR: Egocentric 4D Demonstrations of Human Terrain Traversal

Paper • 2607.13472 • Published Jul 15