policy_IDM_GT_40000 — Stretch serve policy (GR00T N1.6-3B)

Run B: base = policy_IDM_AUG (368ep, IDM_GT labels, 3ep) → continued on GT-only originals 40ep for +20,000 steps (GB64). Purpose: re-anchor GT behavior (finish/wrist) after augmentation-dominant training. Trained 2026-08-02, legacy preprocessing (shortest-edge), absolute actions.

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: legacy shortest-edge 256 (NOT letterbox), absolute actions
• 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)