README.md

---
pretty_name: Eden example assets
tags:
- robotics
- motion-imitation
- humanoid
- dexterous-manipulation
size_categories:
- n<1K
---

# Eden — example assets

Sample data that [Eden](https://github.com/Kashu7100/Eden)'s registered tasks resolve **by
default**, so a task is runnable with no arguments on a pip-installed Eden and not only in a
source checkout ([Eden#998](https://github.com/Kashu7100/Eden/issues/998)).

**Layout convention:** one directory per Eden task, named for the registered task name that
reads it (`eden list`). A dataset class claims its directory with `DatasetBase._hf_subdir`, and
sees only that directory — both the download and the file glob are restricted to it — so this
repo can grow without changing what any existing task trains on.
`tests/test_shared_dataset_repo.py` pins the convention: a claimed directory must be a
registered task name, and no two dataset classes may claim the same one.

| Directory | Read by | Contents |
|---|---|---|
| `beyondmimic/` | `eden.tasks.beyondmimic.dataset.BvhCsvG1Dataset` | one BVH clip retargeted to the Unitree G1 (29 DOF), 60 Hz |
| `contrack/` | `eden.tasks.contrack.dataset.DexterHandDataset` | four xArm7 + XHand1 hand-object reference clips |
| `assistmimic/` | `eden.tasks.assistmimic` (`assistmimic`, `assistmimic_bed`) | two caregiver/recipient pair-clip sets on PHC's SMPL humanoid, 30 Hz |

```python
from eden.tasks import TASK_REGISTRY

TASK_REGISTRY.build("beyondmimic")   # fetches beyondmimic/ on first use
TASK_REGISTRY.build("contrack")      # fetches its default clip from contrack/
TASK_REGISTRY.build("assistmimic")   # fetches assistmimic/interx_cluster1.npz on first use
```

## `beyondmimic/`

`G1_gangnam_style_V01_bvh_60hz.csv` — 1942 rows, 36 columns per row: `base_pos` (3) ·
`base_quat_xyzw` (4) · `dofs_pos` (29), in the DOF order of
`eden.options.robots.unitree_g1.G1Dofs29`.

Row 1 is a retargeting artifact — the root jumps 11.6 cm between rows 1 and 2, which is a
5.8 m/s root velocity at frame 0, the largest in the whole 32 s clip. Reference-state
initialization resets the robot onto that frame *including its velocity*, so trackers launch
and fall with no error raised. Use `frame_range=(1, 1942)` (or `--frame-range 2 1942` for the
example's `csv_to_npz.py`, whose ranges are 1-indexed and inclusive) when the first frame
matters.

Retargeted from the [BeyondMimic](https://github.com/HybridRobotics/whole_body_tracking)
release; please cite the authors and refer to the upstream project for terms of use.

```bibtex
@article{liao2025beyondmimic,
   author = {{Liao}, Qiayuan and {Truong}, Takara E. and {Huang}, Xiaoyu and {Gao}, Yuman and {Tevet}, Guy and {Sreenath}, Koushil and {Liu}, C. Karen},
    title = "{BeyondMimic: From Motion Tracking to Versatile Humanoid Control via Guided Diffusion}",
  journal = {arXiv e-prints},
     year = 2025,
}
```

## `contrack/`

Four clips of an xArm7 + XHand1 rig manipulating a tracked object. The default clip for
`TASK_REGISTRY.build("contrack")` is `dexterhand-Cylinder_00-fps_20-xarm7_xhand-3600_3700.h5`.

| File | Object |
|---|---|
| `dexterhand-Cylinder_00-fps_20-xarm7_xhand-3600_3700.h5` | cylinder |
| `dexterhand-Cuboid_00-fps_20-xarm7_xhand-5880_5940.h5` | cuboid |
| `dexterhand-Cuboid_01-fps_20-xarm7_xhand-7824_7864.h5` | cuboid |
| `dexterhand-Ring_00-fps_20-xarm7_xhand-10240_10340.h5` | ring |

Each `.h5` carries `qpos` (`(frames, 1, 19)`: 7 arm DOFs + 12 hand DOFs), `is_rhand`,
`base_translation`, `fps`, a `contacts/0` group (`is_contact` / `points` over the 15
`contacts/segment_names` hand segments), and an `object_tracks/0` group with the object's mesh
(`vertices`, `faces`) and its per-frame `translations` / `orientations_xyzw`. The task exports
the object mesh from the clip and sizes the episode from its frame count.

Reference clips for Eden's port of [ConTrack](https://github.com/EmptyBlueBox/ConTrack)
(arXiv:2606.03177); please cite the authors and refer to the upstream project for terms of use.

## `assistmimic/`

Paired caregiver/recipient reference clips for Eden's port of
[AssistMimic](https://github.com/YutoShibata07/AssistMimic), on the 24-body SMPL skeleton of
`humanoids/smpl_humanoid/smpl_humanoid_phc.xml` (`eden_assets`), 30 Hz.

| File | Read by | Clips | Source (AssistMimic `sample_data/`) |
|---|---|---|---|
| `interx_cluster1.npz` | `assistmimic` (floor) | 8 Inter-X clips, 1622 frames | `interx_processed_fixed_v9_cluster_ids_1_n_clusters_10.pkl` |
| `hhi_assist_aa_rm.npz` | `assistmimic_bed` (bed) | 10 HHI-Assist clips, 3016 frames | `hhi-assist_processed_v6_AA-RM-wo-FullAssist.pkl` |

Written by Eden's `examples/projects/mimic/assistmimic/scripts/convert_pair_motion.py` (SMPL-X ->
SMPL, PHC's upright-start frame; Inter-X height-fixed to the floor, HHI-Assist kept at the bed
height it was recorded at). Every array is concatenated over clips (`num_frames`, `fps`,
`clips` per clip; `body_names` the 24 SMPL bodies) and has one copy per role,
`caregiver_*` / `recipient_*`: `gts` body positions, `grs` global and `lrs` local rotations
(wxyz), `gvs` / `gavs` linear / angular velocities, `dvs` DOF velocities.

Under the untrained PHC policy (`main.py play`), two Inter-X clips end within ~16 steps against
~71 on average: `G011T006A035R010`, where the caregiver's hands press on the recipient at ~900 N
on average, and `G029T008A035R000`.

Please cite AssistMimic and the source datasets, and refer to the upstream projects for terms
of use: Inter-X (Xu et al., CVPR 2024) and HHI-Assist (Saadatnejad et al., RA-L 2025).

```bibtex
@inproceedings{shibata2026assistmimic,
  title={Learning to Assist: Physics-Grounded Human-Human Control via Multi-Agent Reinforcement Learning},
  author={Shibata, Yuto and Yamazaki, Kashu and Jayanti, Lalit and Aoki, Yoshimitsu and Isogawa, Mariko and Fragkiadaki, Katerina},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}
```