Single-handed pick task ("grab the bottle", RIGHT hand). Trained at the CloudWalk Robotics Lab (CW-RL), 2026-06, on a 314-episode real-robot teleoperation dataset from the zero-wandering curation family at radius 20 (the least-aggressive smoothing → longer windows), for the UNITREE_G1_SONIC embodiment. Released as a reference fine-tune for teams building manipulation policies on the GR00T + SONIC + MuJoCo/G1 stack.
Predecessors:v2 (…-210ep-v2-finetune) — the former production champion (superseded by v10 …-371ep-v10-finetune, 12/12) (checkpoint-20000, grasps from poses where v1 failed) — v4 (…-417ep-v4-finetune) (radius-5, the most aggressive curation, eval vs v2 pending), and v1 (…-105ep-v1-finetune). v5 explores the opposite end of the zero-wandering family from v4: per the dataset name (zero-wandering-smooth-radius-20), wandering segments (drift without progress toward the grasp) were removed and trajectories smoothed, but the radius-20 filter is far less aggressive than v4's radius-5 — yielding 314 longer windows (~161 frames/ep) instead of v4's 417 short ones.
This is a behavior-cloning fine-tune of the full 3B model. Status: trained 2026-06-22, closed-loop eval complete (9/12) — only checkpoint-20000 is published (one checkpoint per repo, lineage storage policy); the strongest runner-up (zero non-convergence), though it did not beat the v2 champion (see Evaluation).
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
Why checkpoint-20000. On the 50,496-frame radius-20 set, 20,000 steps is ≈19.0 epochs — above v2's validated sweet spot (≈15.3 epochs), so overfit is a live risk. Only checkpoint-20000 is published (one checkpoint per repo, lineage storage policy); no lower-epoch rung is retained. 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-20000 was run closed-loop on the physical G1 + SONIC stack and measured against the v2 checkpoint-20000 champion. 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 was placed at 12 fixed positions spanning one half-side of the table (the table is fixed relative to the robot). These are the same 12 positions used for data collection (see the source dataset card). One trial per position (12 trials) gives the grid result. Legend: 🟥 red = knocked the bottle over · 🟧 orange = stuck in an "indecision" loop, never reaching the bottle · 🟩 green = successfully grasped the bottle. The policy was invoked with the task prompt "grab the bottle" — the same prompt used to collect the data.
⚠️ Single-trial methodology. Each of the 12 grid positions is tested once (1 trial/position); repeated placements at the same spot can succeed or fail differently — the fixed-position repeat test, where run, shows this per-position variance. Treat the grid score as a single-pass signal, not a repeated-trial mean.
Conclusion. The strongest alternative — and the only model with zero non-convergence. v5 never gets stuck in an indecision loop (0/12 vs v2's 1/12), making it the most reliable at reaching the bottle, though its success rate (9/12, 75%) and knock-over rate (3/12, 25%) sit just below v2's. The least-aggressive curation (radius 20 → longer windows) produced the cleanest trajectories. It does not beat v2 on raw success, but its zero-non-convergence profile makes it the strongest runner-up — a candidate for the positions where v2 hesitates. v2 remains in production.
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.
Overfit risk. 20,000 steps = ≈19.0 epochs on 50,496 frames (more than v2's 15.3). Whether ≈19.0 epochs overfits is decided by closed-loop eval. Only checkpoint-20000 is published (one checkpoint per repo, lineage storage policy); no lower-epoch rung is retained.
Single published checkpoint. Only checkpoint-20000 is published (one checkpoint per repo, lineage storage policy). There is no intermediate ~15k rung near v2's sweet spot — recovering one would require re-training with a finer save interval.
Radius-20 curation. v5 uses the least-aggressive member of the zero-wandering family (radius 20): wandering removed + trajectories smoothed, but longer windows retained (~161 frames/ep) → 314 episodes. Whether this beats v4's aggressive radius-5 curation or v2's is decided by closed-loop eval.
Grip anticipation (family-wide). Observed in closed-loop eval: the policy performs the full grasp motion — approaching and closing the hand — even when no bottle is on the table, and even with the table removed entirely, 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.
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 noted 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.
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.
DP speedup resampling (wrist-Cartesian, 3 mm/frame, max_K=40) of the merged set; no segment removal — reference baseline. ≈16 ep @20k lands on v2's sweet spot.
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
1@misc{cwrl_gr00t_grab_bottle_v5_2026,
2 title = {GR00T N1.7 Fine-Tune v5 --- Unitree G1 "grab the bottle" (right hand, SONIC WBC, zero-wandering radius-20 curated 314-ep dataset)},
3 author = {{CloudWalk Robotics Lab}},
4 year = {2026},
5 howpublished = {Hugging Face model repository},
6 url = {https://huggingface.co/cloudwalk-research/gr00t-n17-g1-grab-bottle-rh-314ep-v5-finetune}
7}
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.
