Gom-sy/stretch-icra-run-g

Stretch ICRA Run G

Research checkpoint for the Hello Robot Stretch cup-serving policy experiments.

• —Experiment arm: Run G — relaxed pseudo-label gate
• —Base architecture: NVIDIA GR00T N1.6-3B
• —Training data: icra_G_gate_q99_025_115ep (115 episodes)
• —40 GT (ai_challenge_groot_v2)
• —20 GT relight (cafe_stretch_groot_lit)
• —55 IDM pseudo-labelled (cafe_aug60_relabel_v2, gate-passing)
• —Final training step: 86,103 (27 epochs, global batch 64)
• —Final training loss: 0.0048
• —Cameras: head + gripper (dual)
• —Action horizon: 16

Gate

Episodes 60-119 (IDM pseudo-labelled) were filtered by:

1. 1.arm clip [0, 0.52] applied to GT before scoring
2. 2.Gate 1 — per-axis global Pearson corr >= 0.90
3. 3.Gate 2 — 16-frame window NMAE, q99 aggregate <= 0.25

5 episodes dropped: 64, 79, 88, 104, 119 — all failing on w_yaw. Gate 1 removed 64/79/104/119; Gate 2 removed 88 only.

This sits between Run A (no gate, 120 ep) and Run F (q99 <= 0.15, 109 ep) in the gate-strength series: A 120 > G 115 > F 109 > E 105.

Training configuration

Identical to Runs A/D/E/F except for the corpus:

Files

Policy inference weights plus the matching processor / statistics / configuration. DeepSpeed optimizer and RNG states are intentionally excluded.

Serving requirements — these are not optional:

• —letterbox 320 and relative action must be enabled (STRETCH_N16_LETTERBOX=320, STRETCH_N16_RELATIVE=1, STRETCH_N16_MULAW="wrist,gripper")
• —after loading, recompute normalization anchors with use_percentiles=True. The cached min/max in norm_params will blow up the wrist axis otherwise.
• —the training corpus carries an arm clip stamp (stats.json action q01 = 0.0). Evaluations must clip arm to max(x, 0) on both prediction and ground truth, or models trained with the clipped anchor are penalised unfairly.

Intended use

Research and reproducibility only. Validate robot limits, preprocessing, normalization, and emergency-stop behavior before physical deployment.