Datasets:

cloudwalk-research
/

gr00t-g1-grab-bottle-right-hand-radius-20-merged

LeRobot v2.1 dataset for the Unitree G1 right-hand bottle-grab task. This is a merge of two source teleoperation datasets, curated with the zero-wandering pipeline at smoothing half-width 20 and safe_frames 50, then renumbered into one contiguous set of 502 episodes. It is the largest and best-curated set of the grab-bottle lineage and the training data for the v6 GR00T N1.7 fine-tune.

Each surviving clean sub-segment is emitted as its own episode (no action chunk ever spans a cut); the deliberately-uncurated episodes are kept whole.

Sources & how each was treated

source dataset	source eps	treatment	output eps	output frames
`gr00t-g1-grab-bottle-right-hand-105ep-v1`	105	curated (radius 20, safe_frames 50)	270	52,167
`gr00t-g1-grab-bottle-right-hand-worst-positions-empty-115ep-v3` - eps 0-94	95	curated (radius 20, safe_frames 50)	212	63,074
`gr00t-g1-grab-bottle-right-hand-worst-positions-empty-115ep-v3` - eps 95-114	20	included as-is (no removal)	20	4,776
merged total	220		502	120,017

The last 20 episodes (95-114) of the worst-positions dataset are kept as-is on purpose - they are empty scenes (no bottle on the table, no robot movement) that round out the stationary distribution.

Quick facts


Episodes	502
Frames	120,017 @ 50 fps
Format	LeRobot v2.1
Camera	`ego_view` 640×480 (head), H.264 yuv420p, no audio
Robot	Unitree G1 (29-DoF body) + Inspire FTP hands (6 finger joints + 1 constant-zero 7th channel for Dex3-layout compatibility, via InspireFTPGripMapper)
Task instruction	`"grab the bottle"` (single task, RIGHT hand)
State dim	43 (`observation.state`)
Action	`action.motion_token` [64] (latent) + `action.wbc` [43] whole-body
Mean episode length	4.8 s (239 f)
Sources	105ep-v1 (105 ep) + 115ep-v3 (115 ep)
Curation	radius-20, safe_frames 50, grasp-frame preservation
Collection	PICO 4 Ultra teleop (headset + 2 controllers, 3-pt VR) + 2 ankle trackers -> SMPL retargeting (CloudWalk Robotics Lab)
Visibility	Public
License	Apache-2.0

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 (6 finger joints + 1 constant-zero 7th channel for Dex3-layout compatibility, 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).

The worst-positions set additionally targets hard, failure-prone start poses of the right hand (the positions where the v1/v2 policies stalled ~10 cm short of contact).

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.

Field	Shape	Meaning
`observation.state`	43	left_leg 6 + right_leg 6 + waist 3 + left_arm 7 + left_hand 7 + right_arm 7 + right_hand 7
`action.wbc`	43	whole-body joint targets (same layout as `observation.state`)
`action.motion_token`	64	latent action token decoded by the SONIC WBC
`observation.eef_state`	14	left/right wrist pos (3) + abs quat (4), x2
`observation.images.ego_view`	480×640×3	head camera (video)

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

Curation method

Kinematic goal-progress curation (no vision): the goal is the right-wrist position at the grasp instant, a frame is "bad" when the smoothed distance-to-goal rises, and contiguous rising runs are removed. The sensitivity is controlled by the smoothing radius (half-width); here 20, with safe_frames = 50 (minimum kept-block length).

Never removed (two gates):

before the point of maximum distance (hand getting into position)
at or after the grasp (always kept - grasp-frame preservation, an improvement over earlier curation runs that had let frames near the grasp be removed)

Post-processing:

pad each removed block by the smoothing half-width (compensates the blur the smoothing introduces)
drop kept blocks shorter than safe_frames (50 here)

Of the 144,525 frames in the curated portions (105-ep set + 115-ep eps 0-94), 20.3 % (29,284 frames) were removed as non-purposeful wandering; 115,241 curated frames survive. The 4,776 as-is frames are added on top -> 120,017 frames total.

Provenance for every output episode (source dataset, source episode, source window, curation mode) is recorded in meta/episodes.jsonl.

Note on safe_frames. The single-source radius-* variants use safe_frames = 40; this merged set uses safe_frames = 50 on its curated passes, so the 105-ep set yields 270 segments here vs 314 in the standalone radius-20 set - not a discrepancy, a stricter minimum-block threshold.

Reproduction (from the source repo):

python gear_sonic/scripts/merge_curated_datasets.py \
    --ds1 gr00t-g1-grab-bottle-right-hand-105ep-v1 \
    --ds2 gr00t-g1-grab-bottle-right-hand-worst-positions-empty-115ep-v3 \
    --output-path gr00t-g1-grab-bottle-right-hand-radius-20-merged \
    --smooth-radius 20 --safe-frames 50 --ds2-curate-until 95 --write

What this dataset is - and is not

Is: the largest, best-curated grab-bottle set - a radius-20 merge of the 105-ep source and the 115-ep worst-positions source, with grasp-frame preservation. The training data for the v6 GR00T N1.7 fine-tune.
Is not: a single-source curation. The four zero-wandering-smooth-radius-{5,10,15,20} sets curate the 105-ep source alone.
Is not: the un-segmented source sets. Use the 105-ep / 115-ep sources for the un-curated distribution.
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-radius-20-merged",
    repo_type="dataset",
    local_dir="data/real/gr00t-g1-grab-bottle-right-hand-radius-20-merged",
)

How to load

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset(
    "gr00t-g1-grab-bottle-right-hand-radius-20-merged",
    repo_id="cloudwalk-research/gr00t-g1-grab-bottle-right-hand-radius-20-merged",
)
print(ds.num_episodes, ds.num_frames, ds.fps)

How to use for fine-tuning & inference

This set trained the v6 GR00T N1.7 fine-tune (checkpoint-10000, checkpoint-20000) - the largest of the lineage. Fine-tuning was run on NVIDIA B200 GPUs with the Isaac-GR00T stack; see the model card for the exact hyperparameters.

To run the resulting policy in inference (policy server), from an Isaac-GR00T environment:

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 same 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

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.
Curated, not source. The curated portions are clean sub-segments; the 20 worst-positions episodes (95-114) are kept as-is on purpose. "Episode length" mixes clean segments and full episodes.
safe_frames = 50 on the curated passes (vs 40 for the single-source radius-* variants) - see the note in Curation method.
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.
Empty-scene episodes. The 20 as-is episodes (115ep eps 95-114) are empty scenes (no bottle, no movement) kept to round out the stationary distribution - not harder grasps.

Lineage

This set is the merge of the two source sets, both curated at radius-20. Its direct siblings are the four single-source radius-* variants (same method, 105-ep alone) and the speedup-3mm-v1 / speedup-3mm-cycle-removed-v1 sets (same two sources, DP resampling instead of zero-wandering - the latter trains v7). The right-hand-v2 set is a 105-ep-only distance-to-goal curation (the v2 champion). The full family is in the table below.

                  +--> zero-wandering-smooth-radius-{5,10,15,20}   (105ep curated ALONE)
105ep-v1 ---------+--> right-hand-v2                              (105ep; dist-to-goal -> v2 champion)
                  |
                  +--> radius-20-merged (this)   (BOTH curated at radius-20; 502 ep -> v6)
115ep-v3 ---------+--> speedup-3mm-v1            (BOTH; DP speedup resample, NO removal; 220 ep -> v7)
                  +--> speedup-3mm-cycle-removed  (BOTH; DP + cycle removal; 405 ep, no model)

Dataset	Derived from	Episodes	Frames	Notes
105ep-v1	- (source)	105	70,680	Base teleoperation set -> v1 fine-tune
worst-positions-empty-115ep-v3	- (source)	115	78,621	Worst-positions + 20 empty-scene episodes (eps 95-114, no bottle/no movement)
right-hand-v2	105ep-v1	210	62,772	Distance-to-goal wandering removal -> v2 fine-tune (production champion)
zero-wandering-smooth-radius-5	105ep-v1	417	48,577	Zero-wandering r=5 (most aggressive) -> v4 fine-tune
zero-wandering-smooth-radius-10	105ep-v1	371	49,578	Zero-wandering r=10 (experiment, no fine-tune)
zero-wandering-smooth-radius-15	105ep-v1	342	50,634	Zero-wandering r=15 (experiment, no fine-tune)
zero-wandering-smooth-radius-20	105ep-v1	314	50,496	Zero-wandering r=20 (least aggressive) -> v5 fine-tune
radius-20-merged (this)	105ep-v1 + 115ep-v3	502	120,017	Both sources, zero-wandering r=20, grasp-frame preservation; 115ep eps 0-94 curated, 95-114 as-is -> v6 fine-tune
speedup-3mm-v1	105ep-v1 + 115ep-v3	220	60,163	DP wrist-Cartesian resample 3mm/frame, max_K=40, NO segment removal -> v7 fine-tune (trained 2026-06-25, checkpoint-20000 only)
speedup-3mm-cycle-removed-v1	105ep-v1 + 115ep-v3	405	47,944	DP resample + cycle detection + sub-episode splitting (no fine-tune yet)

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 the kinematic zero-wandering pipeline. Stored in the LeRobot v2.1 format.

Citation

@misc{cwrl_grab_bottle_radius_20_merged_2026,
  title        = {GR00T G1 "grab the bottle" --- right-hand radius-20 merged curation (105ep + 115ep sources, 502 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-radius-20-merged}
}

@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.