GR00T N1.7 — G1 + Dex3, 40-step action chunk ablation
This is the weakest of a three-model ablation set and is not meant for use. It is published so
the comparison is reproducible. Use
gr00t-n17-g1-dex3-nubjuki instead — it
is 36 % better.
The ablation set
All three runs share the dataset, the 90/10 split, 10k steps, effective batch 64, 30 fps, 3 cameras, frozen VLM backbone, and a 4×A100 setup. Exactly one thing changes per row. Evaluation is open-loop on the held-out 10 episodes, MAE in radians, using the first 16 predicted steps for all three models so the numbers are directly comparable.
| model | arms | hands | chunk | best ckpt | MAE | left_arm | right_arm | left_hand | right_hand |
|---|---|---|---|---|---|---|---|---|---|
…-nubjuki | RELATIVE | ABSOLUTE | 16 | 10000 | 0.04553 | 0.01700 | 0.04844 | 0.02064 | 0.09605 |
…-nubjuki-abs | ABSOLUTE | ABSOLUTE | 16 | 8000 | 0.06169 | 0.05706 | 0.07386 | 0.02255 | 0.09330 |
| this | ABSOLUTE | ABSOLUTE | 40 | 4000 | 0.07086 | 0.05936 | 0.08322 | 0.02216 | 0.11870 |
Two findings:
- RELATIVE arms beat ABSOLUTE by 26 %. The gap is largest on the left arm (3.4×), which barely moves in this task — a relative model only predicts "no change", an absolute one must regress the exact pose every step. Hands are ABSOLUTE in all rows and score nearly the same, which is the internal control showing the arms caused the difference.
- Extending the chunk from 16 to 40 costs another 15 % (0.06169 → 0.07086), and the optimum moves earlier (4k instead of 8k). Spreading the loss over 40 future steps — the far ones being intrinsically ambiguous — dilutes the near-term precision that actually drives the gripper. GR00T's pretrained embodiments use 40–50 step chunks, but that convention did not transfer to this small-data (17.7 min), precision-grasping task.
Checkpoint sweep (this model)
| checkpoint | MAE | MSE |
|---|---|---|
| 2000 | 0.073154 | 0.020890 |
| 4000 | 0.070859 | 0.021720 |
| 6000 | 0.072146 | 0.020926 |
| 8000 | 0.072401 | 0.021238 |
| 10000 | 0.072481 | 0.021132 |
Scope of the conclusion
The evaluation reads only the first 16 of the 40 predicted steps, i.e. it measures the regime where you re-plan often (execution horizon ≈ 8). A 40-step chunk's potential advantage is a longer execution horizon — useful when inference latency is high — and that regime is not tested here. On the deployment path this project targets (~100–160 ms end-to-end), a 16-step chunk at 30 fps already leaves ample slack, so the longer chunk buys nothing it can trade its accuracy loss for.
Open-loop error is a proxy; closed-loop evaluation on the robot is the real test.
Interface
Predicts absolute 28-d joint targets in radians as a 40-step chunk, intended for 30 Hz execution.
Observation: 3 RGB cameras 480×640 (ego_view ← head, right_wrist_view, left_wrist_view) +
current 28-d joint state + the exact training string
"pick up the doll and place it in the red region."
Joint order (state and action alike): left_arm 0–6, right_arm 7–13, left_hand 14–20,
right_hand 21–27.
⚠️ The hands do not share a finger order: left is Thumb0-2 → Middle0,1 → Index0,1, right is Thumb0-2 → Index0,1 → Middle0,1, and the bending signs are mirrored. This comes from the source dataset's metadata and the model learned it as-is.
Dataset: carroll511/g1_dex3_nubjuki
(100 episodes / 31,803 frames / 30 fps, single task).
Serving requires Isaac-GR00T (validated at 65cc4a192), transformers 4.57.3, torch 2.7.1+cu128, and
access to the gated nvidia/Cosmos-Reason2-2B tokenizer.