cloudwalk-research/gr00t-g1-grab-bottle-right-hand-speedup-3mm-cycle-removed-v1

Grab-Bottle (right hand) - DP-resampled 3 mm/frame + cycle removal

LeRobot v2.1 dataset for the Unitree G1 right-hand bottle-grab task. Extends the DP-resampled variant by automatically detecting and removing return cycles in wrist-space - segments where the arm travels away from a point and then returns to (approximately) the same position. Such cycles teach the policy to backtrack rather than commit to a goal-directed motion. No fine-tune has been produced from this variant yet. A `v8` model card exists as a placeholder — training pending, no weights published.

Each detected cycle is cut out of the episode. The remaining contiguous portions are emitted as independent sub-episodes. The region around the grasp point is always protected.

Quick facts

Collection setup

Same data collection as the 105-ep source set and the worst-positions 115-ep set

• —see those cards for the full setup. In brief:

• —Robot: Unitree G1 (29-DoF body) + Inspire FTP hands (7-DoF/hand: 6 finger joints + 1 grip pad, via InspireFTPGripMapper)
• —Whole-body controller: SONIC WBC
• —Teleoperation: PICO 4 Ultra (headset + 2 controllers, 3-point VR tracking) + 2 ankle trackers -> SMPL retargeting
• —Camera: Intel RealSense D435I on the G1 head; Hand control: grip buttons close the Inspire FTP right hand during the grasp
• —Task prompt: the teleop demonstrations were collected under the task prompt "grab the bottle"; the same prompt is used to invoke the policy at inference (see the model cards' "Prompt conditioning" caveat).

Source datasets

Action and state layout

This is a GR00T + SONIC dataset, not a flat-joint-vector dataset. The policy action is emitted as a 64-D latent `motion_token` (decoded by the SONIC whole-body controller into whole-body joint commands) plus 7-D left/right hand-joint targets - it only produces motion through the SONIC WBC + ZMQ deploy pipeline, not on a bare G1.

See meta/info.json for the full feature schema, including teleop.* retargeting fields and observation.* base state.

Curation method

Step 1 - DP resampling

Identical to the base dataset. Wrist XYZ from observation.eef_state[7:10] (world frame). The DP solver keeps frames where consecutive wrist displacement is ~3 mm/frame.

Step 2 - cycle detection

Operates entirely in DP-frame space (duration is counted in kept frames, not original frames).

For each DP-frame starting point a, find the future DP frame b with the minimum XYZ return distance (epsilon) among all frames at least min_dur kept-frames away. Score the candidate: score = epsilon / dur^k.

Candidates are filtered by hard caps and then greedily selected in score order (best first, non-overlapping).

Score formula: score = epsilon / dur^2

• —Lower = better (tight return + long duration).
• —dur_exp=2 ensures that when a half-cycle and a full cycle overlap, the full cycle wins if its dur^2 advantage outweighs its larger epsilon.

Step 3 - episode splitting

Detected cycles are cut out. The contiguous portions between cycles become separate sub-episodes. Sub-episodes shorter than 40 DP-kept frames are discarded.

Grasp-protection rule: the last sub-episode is never discarded if it contains the grasp frame, regardless of length. The grasp is the reward signal; losing it destroys training value. This rule was triggered 0 times in this run (the min-frames filter discards pre-grasp segments, not post-grasp ones).

What this dataset is - and is not

• —Is: the DP-resample + cycle-removal variant - the DP-only baseline with return-cycle segments cut out and the survivors split into sub-episodes.
• —Is: a curation experiment - no fine-tune has been produced from it yet. The fine-tuned sibling is v7 (from the DP-only baseline).
• —Is not: zero-wandering curated; it replaces wandering detection with cycle detection. It is not a superset of the DP-only baseline (cycles are removed, not retained).
• —Right-hand only, single camera, single task - see caveats below.

How to download

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="cloudwalk-research/gr00t-g1-grab-bottle-right-hand-speedup-3mm-cycle-removed-v1",
    repo_type="dataset",
    local_dir="data/real/gr00t-g1-grab-bottle-right-hand-speedup-3mm-cycle-removed-v1",
)

How to load

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset(
    "gr00t-g1-grab-bottle-right-hand-speedup-3mm-cycle-removed-v1",
    repo_id="cloudwalk-research/gr00t-g1-grab-bottle-right-hand-speedup-3mm-cycle-removed-v1",
)
print(ds.num_episodes, ds.num_frames, ds.fps)

How to use for fine-tuning & inference

No fine-tune has been produced from this variant - it is a curation experiment stacking cycle removal on the DP-only baseline (that baseline trained v7). To fine-tune this variant yourself, use the Isaac-GR00T stack on NVIDIA B200 GPUs, then serve with:

python -m gr00t.eval.run_gr00t_server \
    --model-path <checkpoint-dir> \
    --embodiment-tag UNITREE_G1_SONIC \
    --device cuda:0 --host 0.0.0.0 --port 5550

The server must be launched with the UNITREE_G1_SONIC embodiment tag used in training. Closed-loop control of the G1 is driven by the SONIC WBC in GR00T-WholeBodyControl; see the NVlabs VLA inference tutorial.

Caveats

1. 1.Right-hand-only, single task, single camera. Only the head ego_view was recorded - no wrist cameras. Left-hand and locomotion dims reflect the largely stationary demonstrations.
2. 2.Curated, not source. Episodes are clean sub-segments (cycles cut out), so "episode length" is short (~118 f). Use the sources for the un-segmented distribution.
3. 3.No fine-tune yet. A curation experiment; the DP-only baseline trained v7. A `v8` model card exists as a placeholder (training pending, no weights).
4. 4.Empty-scene episodes kept as-is. ds2 eps 95-114 (no bottle, no movement) are included raw.
5. 5.Deployment needs the SONIC stack. Actions are motion_token + hand joints for the UNITREE_G1_SONIC embodiment; they only produce motion through the SONIC WBC + ZMQ deploy pipeline, not on a bare G1.

Lineage

This set is the DP-resample + cycle-removal variant. Its direct sibling is the DP-only baseline (same DP resampling, no cycle removal -> v7). It shares its two sources with radius-20-merged but uses a different method. The full family is in the table below.

Attribution

Collected at the CloudWalk Robotics Lab (CW-RL) using PICO 4 Ultra teleoperation on the Unitree G1 with the SONIC whole-body controller, then curated with DP resampling + wrist-space cycle removal. Stored in the LeRobot v2.1 format.

Citation

@misc{cwrl_grab_bottle_speedup_3mm_cycle_removed_v1_2026,
  title        = {GR00T G1 "grab the bottle" --- right-hand DP-resampled + cycle-removal curation (405 episodes, SONIC WBC)},
  author       = {{CloudWalk Robotics Lab}},
  year         = {2026},
  howpublished = {Hugging Face dataset repository},
  url          = {https://huggingface.co/datasets/cloudwalk-research/gr00t-g1-grab-bottle-right-hand-speedup-3mm-cycle-removed-v1}
}

@article{gr00t_n1_2025,
  title         = {{GR00T N1}: An Open Foundation Model for Generalist Humanoid Robots},
  author        = {{NVIDIA}},
  year          = {2025},
  eprint        = {2503.14734},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2503.14734}
}

@article{sonic_2025,
  title         = {{SONIC}: Whole-Body Controller for Humanoid Robot Teleoperation},
  year          = {2025},
  eprint        = {2511.07820},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2511.07820}
}

@software{lerobot2024,
  title  = {{LeRobot}: State-of-the-art Machine Learning for Real-World Robotics in {PyTorch}},
  author = {{Hugging Face}},
  url    = {https://github.com/huggingface/lerobot},
  year   = {2024}
}

License

Released under Apache-2.0 by CloudWalk Research. Demonstrations were collected on hardware owned by CloudWalk; the SONIC WBC and GR00T stacks retain their respective upstream licenses. Research preview - closed-loop deployment on a physical humanoid requires human oversight.