Single-handed pick task ("grab the bottle", RIGHT hand). Trained at the CloudWalk Robotics Lab (CW-RL), 2026-06, on the largest dataset in the lineage (371 episodes / 138,546 frames) for the UNITREE_G1_SONIC embodiment. Released as a reference fine-tune for teams building manipulation policies on the GR00T + SONIC + MuJoCo/G1 stack.

What makes v10 different — the largest set, curated by the "Break Down" (v1 wandering-detection) method. v10's dataset is the merge of the two raw sources — gr00t-g1-grab-bottle-right-hand-105ep-v1 (DS1, 105 ep) + …-worst-positions-empty-115ep-v3 (DS2, 115 ep, harder starting positions) — with Break-Down wandering removal: rising segments of the distance-to-goal curve are detected and removed, and each kept sub-segment becomes its own episode (so no GR00T action chunk spans a cut). Params: rise_thresh 24 mm/s, smooth_radius 7, min_window 40, grasp_frac 0.95. Result: 371 episodes (210 DS1 sub-segments + 141 DS2 sub-segments + 20 DS2 as-is) / 138,546 frames (124,077 training samples) — the largest set in the lineage. This makes v10 a curation-method sibling of v6 (same merged DS1+DS2 source; v6 uses radius-20 with grasp-frame preservation, v10 uses Break-Down v1-wandering). The natural comparison is v10 vs the v2 champion — does the largest Break-Down-curated set (at ~10 epochs) beat the champion's curated data? — with v10 vs v6 as the curation-method comparison on the shared merge. It is not step-comparable to v7/v8/v9 (different processing and a different step budget → confounded).

Predecessors: v2 (…-210ep-v2-finetune) — the former production champion (checkpoint-20000, 10/12; superseded by v10, 12/12) — v4 (radius-5, 6/12), v5 (radius-20, 9/12), v6 (merged radius-20, 8/12), v7 (speedup-3mm raw, 7/12), v8 (speedup-3mm + cycle-removed raw, 1/12), v9 (speedup-2mm on curated, 8/12) — and v1 (105ep, 4/12).

This is a behavior-cloning fine-tune of the full 3B model. Status: trained 2026-07-06, closed-loop eval complete — checkpoint-30000 is the new production champion (12/12, beats v2's 10/12) (train loss 0.0280 @ 30k / 0.0396 @ 20k / 0.0676 @ 10k; all three checkpoints published here). Recipe: 30,000 steps (user choice — not the 20k of v5–v9), save every 10,000 → {checkpoint-10000, checkpoint-20000, checkpoint-30000}. See Evaluation.

Quick facts


Base	GR00T-N1.7-3B — Qwen3-VL vision-language backbone + flow-matching diffusion-transformer (DiT) action head
Parameters	3.14 B total / 1.62 B trainable (51.5%)
Dataset	cloudwalk-research/gr00t-g1-grab-bottle-right-hand-v10 — 371 episodes, 138,546 frames @ 50 Hz (124,077 training samples), 480×640 `ego_view` camera (no wrist cams); largest set in the lineage — merged DS1 (105ep) + DS2 (115ep worst-positions), curated by the Break-Down v1-wandering method
Robot target	Unitree G1 (29-DoF body) + built-in G1 hands (7-DoF/hand) + SONIC whole-body controller
Embodiment tag	`UNITREE_G1_SONIC` (`unitree_g1_sonic`)
State space	43-D (left_leg 6 + right_leg 6 + waist 3 + left_arm 7 + left_hand 7 + right_arm 7 + right_hand 7)
Action space	`[40 × 78]` = 40-step horizon × (64 motion_token + 7 left_hand_joints + 7 right_hand_joints)
Hardware	6× NVIDIA B200 (sm_100 / Blackwell)
Mixed precision	bf16
Optimizer	AdamW, lr 1e-4 cosine, warmup_ratio 0.05, weight_decay 1e-5
Steps / batch	30,000 / global batch 48 (8 per GPU × 6 GPUs); checkpoints saved every 10,000 → `{checkpoint-10000, checkpoint-20000, checkpoint-30000}`
Epochs	10.39 @ 30k / 6.93 @ 20k / 3.46 @ 10k (1.44M frame-views ÷ 138,546 frames)
Augmentation	color jitter (brightness 0.3, contrast 0.4, saturation 0.5, hue 0.08)
Wall-clock	1:34:02 for 30k steps on 6× B200; steady-state ~5.3 it/s
Final train loss	0.0280 @ 30k · 0.0396 @ 20k · 0.0676 @ 10k (run min 0.0207, mean 0.0739)
W&B run	`redacted` (project `g1_grab_bottle`; offline → synced)

Repository contents

checkpoint-10000/    # 10k steps (~3.5 ep)
checkpoint-20000/    # 20k steps (~6.9 ep)
checkpoint-30000/    # 30k steps (~10.4 ep) — most-trained rung, natural production candidate
README.md            # this file

All three {checkpoint-10000, checkpoint-20000, checkpoint-30000} are published so the closed-loop eval can pick the best one. Because the dataset is the largest in the lineage, all three rungs sit below v2's validated ~15.3-epoch sweet spot — 30k (~10.4 ep) is the most-trained and the default bet.

Each checkpoint-NNNNN/ is a self-contained, deploy-ready snapshot:

model-00001-of-00002.safetensors + model-00002-of-00002.safetensors (~6.5 GB, bf16)
model.safetensors.index.json
config.json
embodiment_id.json (contains unitree_g1_sonic)
processor_config.json
statistics.json (dataset normalization stats)
experiment_cfg/ (training config snapshot)

DeepSpeed ZeRO partition states, optimizer, scheduler and RNG files are intentionally omitted — they are needed only to resume training and would add several GB per checkpoint.

Evaluation

Epoch math. This run trains for 30,000 steps = 10.39 epochs on the 138,546-frame set (20,000 = 6.93; 10,000 = 3.46). The set is the largest in the lineage, so even 30k stays below v2's validated ~15.3-epoch sweet spot. At the 20k baseline of v5–v9 this set would be the most underfit run (6.9 ep, below v6's 8.0); 30k was chosen (vs the locked 20k of v5–v9) to push the epoch count up without a very long run; 30k is the natural production rung, with 10k/20k showing the trajectory. The production checkpoint is chosen by closed-loop comparison, not by training loss. There is no held-out split; train loss is a fit probe, not a generalization measure.

Closed-loop evaluation — completed (real G1). checkpoint-{10000,20000,30000} were run closed-loop on the physical G1 + SONIC stack and measured against each other and the v2 checkpoint-20000 champion (10/12 across hand-placed bottle poses). Low BC loss is a healthy prerequisite, not evidence of a good policy — this closed-loop result is the load-bearing signal.

Procedure. The bottle is hand-placed at 12 fixed grid positions on the table (spreading near / far / lateral reaches), 1 trial per position, right-hand grab. Outcome per position:

🟩 success — bottle grabbed and lifted cleanly
🟧 non-convergence — policy loops / indecision, never completes the grab within the trial
🟥 knock-over — bottle knocked over (or off the table) before a clean grab

Result — checkpoint-30000: 12/12 (100%) 🟩 — new production champion. Every position grabbed cleanly on the first trial. This beats the v2 champion (10/12, 83% — 1 🟥 knock-over + 1 🟧 non-convergence) and is the best closed-loop result in the lineage.

Closed-loop eval: 12-position grid, checkpoint-30000

⚠️ Single-trial methodology. The 12/12 reflects 1 trial per position. A same-position repeat test (the fixed "black-rectangle" position, 5 trials) landed 4/5 (80%) — i.e. repeated placements at the same spot can fail, and the per-position success rate under repeated sampling is lower than the 1-of-1 grid suggests. Treat the 12/12 as a strong single-pass signal, not a repeated-trial mean. (The 10k and 20k rungs were also tested: results were similar to 30k, with ~1 extra 🟥 knock-over each; the 30k rung was kept as the recommended production checkpoint — no reason found to prefer an earlier rung.)

Qualitative observations. On the far positions the policy takes visibly longer to commit ("thinks about it") but still converges and completes the grab. On the near positions it makes a notably clean elbow adjustment — a motion quality attributed to the extra DS2 (worst-positions) episodes in the merged training set, which broadened the starting-pose distribution.

Conclusion. v10 is the new production champion. The win confirms the thesis: the v2-style distance-to-goal curation ("Break Down"), applied to the larger 105+115 merge, scales — more episodes under the same curation that made v2 the champion beats v2 itself. This is the family's first rung to clear 12/12. Notably it does so at only ~10.4 epochs (well below v2's ~15.3-epoch sweet spot), so the gain comes from the data (more curated episodes), not from longer training. The trade-off seen across the speedup family (committed motion vs close-position precision) does not appear here — v10 keeps the near-position adjustment quality while still converging on the far positions. checkpoint-30000 enters production; v2 is superseded.

How to download

python

from huggingface_hub import snapshot_download

local = snapshot_download(
    repo_id="cloudwalk-research/gr00t-n17-g1-grab-bottle-rh-371ep-v10-finetune",
    repo_type="model",
    allow_patterns=["checkpoint-30000/*"],  # or checkpoint-10000/* / checkpoint-20000/*
)
print(local)

How to deploy

Start the GR00T policy server (from an Isaac-GR00T environment) pointing at the downloaded checkpoint:

bash

python -m gr00t.eval.run_gr00t_server \
    --model-path <local>/checkpoint-30000 \
    --embodiment-tag UNITREE_G1_SONIC \
    --device cuda:0 --host 0.0.0.0 --port 5550

Closed-loop control of the G1 (sim or real) is driven by the SONIC whole-body controller in GR00T-WholeBodyControl: the policy emits motion_token + hand-joint targets that the SONIC WBC decodes into whole-body joint commands. The server must be launched with the same UNITREE_G1_SONIC embodiment tag used in training. See the NVlabs VLA inference tutorial. This checkpoint is not plug-and-play on hardware — it requires the SONIC C++ deploy stack and the matching G1 setup.

Known caveats

Right-hand-only, single task. The dataset is one task ("grab the bottle") executed with the right hand. Left-hand and locomotion action dims reflect the (largely stationary) demonstrations; do not expect bimanual or locomotion behavior.
Single camera. Only the ego_view (head) camera was recorded — no wrist cameras. The policy has never seen wrist views.
Deployment needs the SONIC stack. The checkpoint outputs motion_token + hand joints for UNITREE_G1_SONIC; it only produces robot motion through the SONIC WBC + ZMQ deploy pipeline. It is not directly executable on a bare G1.
Train loss is not held-out. No episode split; the loss is a smoothed train-fit probe, not a generalization metric.
Largest set, Break-Down-curated. v10's data is the biggest in the lineage (138,546 frames) — the merged DS1+DS2 set curated by the Break-Down v1-wandering method (a curation-method sibling of v6). The risk is dilution: 138,546 frames at 30k steps is only ~10.4 epochs. Compare against the v2 champion (and v6 for the curation-method angle), not against v7/v8/v9 as a processing-isolation pair (different processing + step budget).
Epoch zone — underfit-leaning. 30,000 steps ≈ 10.4 epochs; 20k ≈ 6.9; 10k ≈ 3.5. All three rungs sit below v2's ~15.3-epoch sweet spot (the most underfit run of the lineage). The eval — not the training loss — picks the rung; 30k is the default bet, and a longer extension may be warranted if even 30k underfits.
Three published checkpoints. checkpoint-10000, checkpoint-20000 and checkpoint-30000 are all published so the closed-loop eval can pick the best rung.
Training environment. Trained on 6× B200 GPUs with NCCL_IB_DISABLE=1 NCCL_P2P_LEVEL=NVL (InfiniBand off, P2P over NVLink) and W&B in offline mode (synced post-run). These affect only the training run, not the weights.
Grip anticipation (family-wide). Observed in closed-loop eval across the family: the policy performs the full grasp motion — approaching and closing the hand — even when no bottle is on the table, as if the bottle were present. This grip-anticipation artifact is shared across the whole grab-bottle family (the same training demonstrations). Worth gating on bottle detection for downstream deployment. (Directly observed on v10 too: with no bottle on the table — and even with the table removed entirely — the policy still performs the full grasp motion, approaching and closing the hand as if the bottle were present.)
Prompt conditioning — untested. All training data was collected with the task prompt "grab the bottle", and the same prompt is used to invoke the policy at inference — a constant across every episode and every model in this family. We have not tested whether conditioning the prompt on the bottle's presence (e.g. "grab the bottle if there is one in the image") would gate the grasp and reduce the grip-anticipation artifact observed above (the robot closing its hand even with no bottle present). It is a plausible mitigation, but unverified: the model may have learned the grasp as an unconditional reflex from the demonstrations, independent of prompt phrasing.

Lineage

Version	Dataset	Episodes	Frames	Epochs	Closed-loop	Notes
v1	105ep-v1	105	70,680	13.6 @20k	4/12 (33%)	First GR00T N1.7 + SONIC fine-tune at CW-RL; right-hand bottle pick.
v2	right-hand-v2	210	62,772	15.3 @20k	✅ validated (10/12, 83%)	Curated (windows split, bad segments removed). `checkpoint-20000` = former production champion (10/12), superseded by v10 (12/12).
v4	radius-5	417	48,577	19.8 @20k	6/12 (50%)	Zero-wandering, most aggressive curation (radius 5).
v5	radius-20	314	50,496	19.0 @20k	9/12 (75%)	Zero-wandering, least aggressive curation (radius 20).
v6	radius-20-merged	502	120,017	8.0 @20k	8/12 (67%)	Merged (105-ep + 115-ep "worst-positions"), radius-20 with grasp-frame preservation.
v7	speedup-3mm-v1	220	60,163	16.0 @20k	7/12 (58%)	DP speedup (wrist-Cartesian, 3 mm/frame) of the raw merged set; no segment removal — raw-branch baseline.
v8	speedup-3mm-cycle-removed-v1	405	47,944	20.0 @20k	1/12 (8%)	DP speedup + segment removal on the raw merged set (the v7 follow-up).
v9	speedup-2mm-v3	193	32,786	29.3 @20k	8/12 (67%)	*Gentle 2 mm DP speedup on the curated* (v2-lineage) data**. Smallest set → 10k ≈ sweet spot; trained (loss 0.0308 @20k).
v10 (this)	right-hand-v10	371	138,546	10.4 @30k	✅ 12/12 (100%) — new champion	Largest set — merged DS1+DS2, Break-Down v1-wandering curation (same curation family as v2, on the merge). 30k steps → ~10.4 ep (below v2's ~15.3 sweet spot); the win comes from more curated data, not longer training. `checkpoint-30000` = new production champion (supersedes v2).

References

GR00T N1 — Open foundation model for generalist humanoid robots; base policy fine-tuned here. Paper (arXiv:2503.14734), base model.
Isaac-GR00T — Training / inference / deployment stack used for this fine-tune. GitHub.
GR00T-WholeBodyControl (SONIC) — Whole-body controller + teleop + VLA deployment for the G1. GitHub, SONIC paper (arXiv:2511.07820).
Training dataset — cloudwalk-research/gr00t-g1-grab-bottle-right-hand-v10 (CloudWalk Research, 2026), PICO 4 Ultra teleoperation on the Unitree G1 with SONIC WBC; largest set in the lineage (371 ep / 138,546 frames) — merged DS1 (105ep) + DS2 (115ep worst-positions), Break-Down v1-wandering curation.

Attribution

Developed by cloudwalk-research in the CloudWalk Robotics Lab (CW-RL). Fine-tuned from nvidia/GR00T-N1.7-3B using Isaac-GR00T; targets the Unitree G1 with the SONIC whole-body controller.

Citation

bibtex

@misc{cwrl_gr00t_grab_bottle_v10_2026,
  title        = {GR00T N1.7 Fine-Tune v10 --- Unitree G1 "grab the bottle" (right hand, SONIC WBC, largest set, Break-Down-curated merge, 30k steps)},
  author       = {{CloudWalk Robotics Lab}},
  year         = {2026},
  howpublished = {Hugging Face model repository},
  url          = {https://huggingface.co/cloudwalk-research/gr00t-n17-g1-grab-bottle-rh-371ep-v10-finetune}
}

License

Inherits the NVIDIA Open Model License Agreement of the base model nvidia/GR00T-N1.7-3B — see the license terms. This is a research preview: not intended for safety-critical use; closed-loop deployment on a physical humanoid requires human oversight.