robotwin-hf / README.md
ethenj's picture
Describe250Hz collection and manualdownloadapproval
e7b9310 verified
|
Raw History Blame Contribute Delete
11.1 kB
---
license: mit
language:
- en
tags:
- robotics
- robotwin
- bimanual-manipulation
- proprioception
- joint-velocity
- 250hz
pretty_name: robotwin-hf
extra_gated_heading: Request access to robotwin-hf
extra_gated_description: Download access requires manual approval by the dataset owner.
size_categories:
- 10K<n<100K
---
# robotwin-hf
**50 RoboTwin tasks · 27,500 successful expert episodes · 250 Hz dense proprioception
and actuator signals · synchronized four-view RGB videos.**
This HMRC collection uses RoboTwin 2.0 with the `aloha-agilex` embodiment and the
`native_joint_velocity_gripper_position_v2` control protocol. Arm commands are
native joint velocities, with zero arm position stiffness; grippers use position targets.
The original compressed **NPZ + FFmpeg MP4** files are preserved byte for byte.
中文:50个任务,每任务50条clean+500条random,包含完整250 Hz关节状态、控制命令和力信息,
以及同步RGB。按任务与来源分包;文末说明状态/动作对齐,避免训推时错配。
## Contents
| Property | Value |
|---|---:|
| Tasks | 50 |
| Clean / random episodes | 2,500 / 25,000 |
| Train / validation / test episodes | 21,842 / 2,885 / 2,773 |
| Synchronized visual observations | 11,243,243, each with four cameras |
| Physics ticks, excluding initial rows | 159,851,744 |
| Physics rate | 250 Hz, 0.004 s/tick |
| RGB resolution | 240 × 320, RGB |
| Cameras | head, left wrist, right wrist, third view |
| Raw receipt-referenced artifact bytes | 174,658,881,096 (~162.66 GiB) |
| Archive files | 100, one per task and source |
Regular visual observations are 15 physics ticks apart (about16.67 Hz), with
additional boundary observations. **MP4 playback is30 FPS; use recorded physics
markers and simulation timestamps for alignment.** Dense streams include every
non-rendered physics tick. Stationary cleanup is disabled, and five visual frames
are recorded after success while expert motion continues.
`hmrc_clean_50` and `hmrc_random_500` describe collection conditions, not train/test
splits. Random collection varies backgrounds, clutter, table height and lighting.
`SHA256(task:seed)[:8] %10` selects test(0), validation(1), train(2–9), keeping
the same task/seed across sources in one split. Task IDs are fixed in
`metadata/tasks.csv` and `metadata/source_catalog.json.gz`.
## Download and extract
The repository page is public. **Dataset file downloads require manual approval.**
Visit this dataset page, submit an access request, wait for approval, and authenticate
your environment with `hf auth login` before using the commands below.
Install FFmpeg/ffprobe and the Python dependencies, then download this repository's
small code/metadata files:
```bash
python -m pip install numpy torch av huggingface_hub
python - <<'PY'
from huggingface_hub import snapshot_download
snapshot_download('ethenj/robotwin-hf', repo_type='dataset', local_dir='robotwin-hf',
ignore_patterns=['archives/**'])
PY
cd robotwin-hf
# One task, both sources (~several GB):
python download.py --task adjust_bottle --destination data
# Or all 27,500 episodes:
python download.py --destination data
```
`download.py` checks each downloaded archive against `manifest.json`, extracts
only regular files within the destination, and preserves the original directory
layout. Downloading a subset is supported with `--source clean` or `--source random`.
Storage must accommodate both archives and extracted data (roughly twice the dataset
size for a full download). `SHA256SUMS` can also be used with `sha256sum -c`.
```text
archives/hmrc_clean_50/<task>.tar
archives/hmrc_random_500/<task>.tar
metadata/{tasks.csv,episodes.csv,summary.json,schema.json,source_catalog.json.gz,...}
manifest.json
SHA256SUMS
command_protocol.json
dataloader/
scripts/
```
Each tar extracts to:
```text
joint_velocity_v2/<source>/<task>/aloha_agilex/
meta/ # visual-row NPZ and camera/action manifest
dense/ # full-rate numeric NPZ and static joint/force metadata
video/ # four MP4 files per episode
instruction/ # episode-specific seen/unseen instructions
_traj_data/ # original expert trajectory pickle
success/ # successful publication receipts and original artifact hashes
command_protocol.json
seed.txt
scene_info.json
```
Logs, failed/rejected attempts, staging data and training checkpoints are excluded.
Tars are not recompressed because NPZ and MP4 are already compressed. The complete
release is indicated by `RELEASE_COMPLETE.json`, published after remote checksums,
an authorized download round-trip, and confirmation that anonymous archive access is
denied. Its absence means upload/verification is still in progress.
## Fast dataloader
Rebuild a local catalog after extraction; `source_catalog.json.gz` describes the original
snapshot, and its source-machine mtime signatures should not be reused for extracted files.
```bash
# For all100 archives; append --allow-partial if only a subset was downloaded.
python build_catalog.py --root data/joint_velocity_v2 --output data/catalog
```
```python
from dataloader import RoboTwinNPZMP4Dataset, make_robotwin_loader
dataset = RoboTwinNPZMP4Dataset(
'data/joint_velocity_v2', catalog='data/catalog/index.json', split='train',
future_frames=32,
cameras=['head_camera', 'left_wrist_camera', 'right_wrist_camera'],
video_backend='pyav', verification='stat')
loader = make_robotwin_loader(dataset, batch_size=4, num_workers=4,
prefetch_factor=2, multiprocessing_context='fork')
for epoch in range(10):
loader.batch_sampler.set_epoch(epoch)
for batch in loader:
# images[camera]: [B,33,H,W,3] RGBuint8; current+32future frames
# dense[field]: [B,max_dense_length,...], no resampling
# dense_mask: valid dense rows; frame_dense_indices: exact image->dense links
pass
```
`fork` is the tested Linux throughput configuration; `spawn` is also supported by
functional tests. Use a short project-local runtime path for multiprocessing sockets.
The factory defaults to persistent workers, bounded caches and a120-second batch timeout.
For two-rank DDP pass `rank`, `world_size=2`, `drop_last=True`, and the same seed.
Episode-local shuffled blocks improve locality; this is not global per-sample shuffle.
The default single-rank sampler covers all windows. Distributed equal-size batching
drops small per-episode tails and blocks that cannot be evenly divided across ranks.
`verification='stat'` checks receipt contracts/content and source size+mtime; it does
not rehash every large file. The constructor default is`hash_once` (full hashes once
per worker/signature), while`hash` rechecks on every episode-cache miss. For a full
local content audit use catalog CLI`--hashes`. The helper also retains all50global
taskIDs for subset downloads. All archive members were checked
against their original receipt SHA256 during packaging.
Measured on the release machine, using three native-resolution33-frame streams and
all dense numeric channels: **28.6 samples/s with4workers** over1,024samples;
**49.8 samples/s aggregate with two concurrent4worker loaders**. These use warm
filesystem caches and stat verification; they are CPU loader measurements, not
end-to-end model training speed. `metadata/loader_benchmark.json` contains details.
## State, action and force semantics
* `state_*_arm_joint_states` and `state_*_ee_joint_states` are measured states;
gripper state is normalized from the actual master-joint qpos, not its command.
Gripper normalization is unclipped and uses the embodiment range[-0.01,0.045]m.
`command_*` holds commands; `state_next_*_ee_poses` holds future observed labels,
which must not be used as current policy inputs.
* Dense row0 is the initial state. Dense row n is the state **after** tick n;
the command in row n acts over(n−1,n]. For next-tick supervision, pair
state[n−1] with velocity-command[n]. A same-row pair would be temporally wrong.
* The arm action is the actual12-joint native velocity target(rad/s); the two
gripper actions are position targets. `robot_drive_target` is a teacher-only
position reference under zero arm stiffness, not an executed arm position action.
Visual-row `action_*` instead stores the command at the next visual observation;
`action_source_frame_index` and `action_physics_step` make that offset explicit.
* Original38-joint dense arrays are preserved; `meta_json` lists the16 useful
joints(12arm joints and4physical gripper joints) and the joint ordering/units.
See`metadata/schema.json` for example shapes/dtypes and names.
* `joint_solver_force_native` is PhysX solver generalized effort: drive, joint
friction and joint-limit terms. It is neither an isolated motor sensor nor an
end-effector wrench. `qf_applied` includes explicit applied/compensation forces;
`total_generalized_force_native` is their sum. Gravity, Coriolis and estimated
PD-drive forces are recorded separately. Revolute units areN·m; prismatic unitsN.
Initial native force is invalid (`native_force_valid=False`).
* Use`dense_mask`,`frame_valid`,`action_valid` and`native_force_valid` for their
respective losses. Dense windows include both endpoint ticks. Some visual
boundaries have repeated/short markers, so a33-image window need not be481ticks.
* Recompute normalization using only the selected training set. No model-specific
normalization from an earlier incomplete collection is provided.
## Validation and provenance
All27,500 episodes have successful atomic receipts and the same250Hz native-velocity
protocol. During this release, **all receipt-referenced artifact content is rehashed
against its original receipt while writing the archives**, and each archive receives
its ownSHA256. The preceding deep audit additionally revalidated100episodes covering
all50tasks and both sources, including numerical/control/time contracts and full
decodes of400videos, with zero failures. The actual-gripper-state check on the same
100episodes agrees with dense qpos within5.96e-8. Six real dataloader windows were
matched exactly byte-for-byte against the older reader;16regression tests passed.
These are data/reader checks, not a learned policy success rate.
Based on [RoboTwin2.0](https://github.com/RoboTwin-Platform/RoboTwin) and the
[HMRC RoboTwin fork](https://github.com/HMRC-Manip/RoboTwin).
The upstream code and original RoboTwin dataset use MIT; the upstream notice is
retained in`LICENSE`. This release is also distributed under MIT. Third-party
components retain their original attribution and terms.
```bibtex
@article{chen2025robotwin,
title={Robotwin2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation},
author={Chen, Tianxing and Chen, Zanxin and Chen, Baijun and Cai, Zijian and Liu, Yibin and Li, Zixuan and Liang, Qiwei and Lin, Xianliang and Ge, Yiheng and Gu, Zhenyu and others},
journal={arXiv preprint arXiv:2506.18088},
year={2025}
}
```