ICRA-S

GR00T-N1.6-3B policy fine-tuned on a Stretch cafe-serving corpus with dual-camera observations (head + gripper).

Training

Base model	GR00T-N1.6-3B
Embodiment tag	NEW_EMBODIMENT
Corpus	120 episodes — 40 GT + 20 relit + 60 pseudo-labeled
Frames	214,566
Epochs	27
Global batch size	64
Steps	90,520
Hardware	2 x A100-PCIE-40GB
Wall time	31 h 20 m
Final loss	0.0060 (mean of last 50 steps)

Tuned modules: projector + diffusion head. Vision tower and LLM stay frozen.

Preprocessing

• Letterbox padding to 320 x 320
• Relative action representation, normalized with relative_stats.json
• Mu-law companding (mu = 3) on the wrist and gripper dimensions
• q99 percentile anchors for normalization
• Arm dimension clipped at q01 = 0.0

Pseudo-labels

The 60 pseudo-labeled episodes come from an inverse dynamics model trained on dual-camera serve data with mu-law (mu = 3). Episodes were admitted by a gate combining a global correlation threshold and a windowed NMAE threshold.

Contents

Weights and configuration only. Optimizer state (global_step90520/) is not included, so this checkpoint is for inference and evaluation rather than resuming training.