CK-Sung/gr00t-n17-g1-dex3-nubjuki-left-abs
GR00T N1.7 — G1 + Dex3 left-hand, all-ABSOLUTE action ablation
This is an ablation control, not the recommended model. Use `gr00t-n17-g1-dex3-nubjuki-left-rel` instead — it is 23 % better. This model is published so the comparison is reproducible.
Everything matches that model except the action representation: same dataset, same 47/6 split, same 10k steps, same effective batch 64, same 16-step chunk at 30 fps, same 3 cameras, same 4×A100 setup. Only left_arm / right_arm changed from RELATIVE to ABSOLUTE (both hands were already ABSOLUTE).
- Dataset: `carroll511/g1_dex3_nubjuki_left_53ep_260730` — 53 episodes / 13,041 frames / 30 fps (7.25 min), single task
- Checkpoint: step 10,000 (best of 2k–10k on val)
Result: RELATIVE arms win
Val open-loop MAE in radians (both models return absolute joint targets after decoding, so the numbers are directly comparable).
The hands are ABSOLUTE in both configurations and score nearly identically — the internal control showing the gap comes from the arms alone. The same experiment on the right-hand dataset (`…-nubjuki` vs `…-nubjuki-abs`) gave RELATIVE a 26 % win with the same per-group signature, so this replicates across two independent datasets.
Checkpoint sweep (this model)
The sweep is nearly flat from 4k onward (0.0938 → 0.0931), i.e. the absolute parameterisation plateaus early on 7.25 minutes of data, whereas the relative model kept improving to 10k.
Interface
Predicts absolute 28-d joint targets in radians, 16-step chunk for 30 Hz execution. Observation: 3 RGB cameras 480×640 (ego_view, right_wrist_view, left_wrist_view) + current 28-d joint state + the task string "pick up the doll and place it in the red region".
Joint order: 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, with mirrored bending signs. This comes from the source dataset's metadata and the model learned it as-is.
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. Open-loop error is a proxy; closed-loop evaluation on the robot is the real test.