Activate_GT_AUG — Stretch serve policy (GR00T N1.6-3B)
Run A: fresh N1.6-3B trained 27 epochs (90,705 steps, GB64, final loss 0.0045) on 120ep 1:1 corpus — GT 60ep (originals 40 + relight 20) with GT labels + repaired augmentation 60ep (left-cup mug prompt, panel excluded, uniform-random targets/colors/materials; labels from GT-trained IDM). Trained 2026-08-03 on gpu3.
Serving requirements (critical)
- Normalization: mu3 —
STRETCH_N16_MULAW="wrist,gripper" (not stored in ckpt)
- After loading, recompute norm anchors:
processor.state_action_processor.use_percentiles=True + _compute_normalization_parameters() (else wrist decodes on min/max anchors and breaks)
- Preprocessing: LETTERBOX 320×320 (aspect-preserving zero-pad) — set
STRETCH_N16_LETTERBOX=320
- Actions: RELATIVE (
STRETCH_N16_RELATIVE=1, dataset modality absolute:false) — server must add current state to decode absolute targets
- Cameras: dual (gripper wrist + head), task text: "serve the cup"
- Robot adapter:
stretch_n16/serve_policy_n16.{py,sh} + gr00t_stretch_policy.py (see README_deploy_n16.md)
