cloudwalk-research/gr00t-g1-grab-bottle-right-hand-worst-positions-empty-115ep-v3

GR00T G1 Grab-Bottle — right hand, worst-positions, 115 episodes

LeRobot v2.1 dataset for the Unitree G1 right-hand bottle-grab task. This is a hard-pose teleoperation set — 115 episodes collected deliberately at the bottle positions where prior fine-tuned policies performed worst (the model would approach the bottle but stop ~10 cm short of contact and fail to grasp). It was recorded to fill the policy's reach-distance distribution gap. Manually cleaned at the episode level; no automated frame-level curation applied.

Together with the 105-ep source set, this set is a source input to the `radius-20-merged` dataset (502 ep / 120,017 frames), which trains the v6 GR00T N1.7 fine-tune. In that merge, episodes 0–94 are zero-wandering curated and episodes 95–114 are kept as-is (static empty-scene episodes — no bottle on the table, no robot movement — that round out the stationary distribution).

Quick facts

Note on naming. The empty token in the dataset name (worst-positions-empty-115ep-v3) marks the 20 empty-scene episodes (eps 95–114): the table has no bottle and the robot does not move. The worst-positions part is the other 95 episodes (eps 0–94), which target the bottle placements where prior models failed. In downstream merges these two groups are treated separately — eps 0–94 are curated, eps 95–114 (the empty scenes) are kept as-is.

Why this set exists

Earlier fine-tunes (notably the v2 champion) converged to an approach pose but did not always push the final ~10 cm to contact. This set adds demonstrations at exactly those hard reaches so the policy sees enough contact-distance examples to close the gap. It is a distribution-filling set, not a general-purpose re-collection.

Curation (manual, episode-level)

Recorded as a single 144-episode session, then cleaned in two passes (episode deletion + reindex; no frame-level editing):

Survivor array columns (joint states, actions, SMPL, video) are byte-identical to the raw session — only episode_index / index were rewritten and bad episodes removed. episodes_stats.jsonl is present for normalization.

Round-2 moved-out episodes (original indices): 6, 8, 13, 65, 67, 75, 84, 85, 89, 92, 95, 126. A full 144-episode backup is retained off-dataset as a safety net.

Collection setup

• —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
• —Operator pose: stationary manipulation (no walking in the recorded segments)
• —Camera: Intel RealSense D435I on the G1 head, streamed to the workstation
• —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).

[image]

Collection grid. The bottle was placed at 12 fixed positions spanning one half-side of the table (the table is fixed relative to the robot); this "worst-positions" set deliberately targets the hard poses on that grid. These are the same 12 positions reused for the closed-loop evaluation of the fine-tuned models (see e.g. the v6 model card), so training and eval share one geometry.

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 (SMPL joints/pose, VR 3-point pose, planner state) and observation.* base state (root orientation, projected gravity, init base quat).

What this dataset is — and is not

• —Is: a hard-pose source set, episode-level curated. Deliberately biased toward the worst-performing bottle placements.
• —Is not: zero-wandering curated. The radius-20-merged set applies the automated kinematic curation (eps 0–94) on top of this.
• —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-worst-positions-empty-115ep-v3",
    repo_type="dataset",
    local_dir="data/real/gr00t-g1-grab-bottle-right-hand-worst-positions-empty-115ep-v3",
)

How to load

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset(
    "gr00t-g1-grab-bottle-right-hand-worst-positions-empty-115ep-v3",
    repo_id="cloudwalk-research/gr00t-g1-grab-bottle-right-hand-worst-positions-empty-115ep-v3",
)
print(ds.num_episodes, ds.num_frames, ds.fps)

How to use for fine-tuning & inference

This set (merged with the 105-ep set, radius-20 curated) trained the v6 GR00T N1.7 fine-tune (checkpoint-10000, checkpoint-20000). 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

1. 1.Distribution-biased. Collected at hard reaches on purpose; not a uniform sample of the task space. Combine with the 105-ep set (or use radius-20-merged) for balanced training.
2. 2.Right-hand-only, single task. Left-hand and locomotion dims reflect the largely stationary demonstrations.
3. 3.Single camera. Only the head ego_view was recorded — no wrist cameras.
4. 4.Not zero-wandering curated. Use radius-20-merged for the frame-cleaned version.
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.
6. 6.`empty` name token — see the note above: the 20 empty-scene episodes (eps 95–114, no bottle/no movement).

Lineage

As a source set, this dataset's direct descendants are the both-source curations: radius-20-merged (→ v6), speedup-3mm-v1 (→ v7), and speedup-3mm-cycle-removed-v1. In each, eps 0–94 (worst-positions) are processed and eps 95–114 (empty scenes) kept as-is. 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. Stored in the LeRobot v2.1 format.

Citation

@misc{cwrl_grab_bottle_worst_positions_115ep_v3_2026,
  title        = {GR00T G1 "grab the bottle" --- right-hand worst-positions teleoperation set (115 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-worst-positions-empty-115ep-v3}
}

@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.