omkarpatil/move-soft-toy-left-groot-sharednorm

GR00T N1.7 — move-soft-toy-left (shared-norm)

Fine-tune of nvidia/GR00T-N1.7-3B on move-soft-toy-left for the ROBOTIS FFW SG2 Rev1, trained with the shared-norm recipe.

Composition group C

Members: move-soft-toy-left, move-soft-toy-right. Normalization statistics were pooled over 5,249 frames of all group members and written identically into each member's dataset, so every policy in the group applies the same invertible transform and their scores can be composed.

sha256(new_embodiment stats)[:16] = a9a2b7939222c30e

Only compose models reporting this same hash. Other groups (push / pick-handover / soft-toy) each have their own transform and are not mutually composable.

Training — fully stock

No code patches. Only the dataset statistics differ from a default fine-tune.

Entrypoint	gr00t/experiment/launch_finetune.py (unmodified)
Normalization	q01/q99 min-max → [-1, 1], use_percentiles=True, clip_outliers=True
Precision	fp32 (load_bf16=False, stock)
Embodiment	NEW_EMBODIMENT — FFW SG2 Rev1, state 22 / action 16
Cameras	cam_left_head, cam_left_wrist, cam_right_wrist
Data rate	15 fps, 16-step action chunk (≈1.07 s)
Episodes / frames	20 / 2575
Steps	20 000, lr 1e-4, warmup 0.05, weight decay 1e-5, batch 32
Final train loss	0.03824643007721752

Attention uses PyTorch sdpa, not flash-attention-2 (host glibc predates the wheel's requirement). Both are exact attention; results are not bit-reproducible against a flash-attn build.

Inference-ready only — optimizer state and intermediate checkpoints are excluded, so this cannot resume training.