surgdex-needlereach-e8

GR00T-H-N1.7 fine-tuned for the SurRoL needle_reach task (dexterous dVRK/PSM, vision-based closed-loop) as part of the SurgDex project.

This is the E8a checkpoint: fine-tuned on a zero-action-tail-trimmed LeRobot dataset with min-max action normalization, latent aux-head disabled, 15,000 steps.

Evaluation (closed-loop SurRoL needle_reach, H=8 receding horizon, denoise=16, N=25)

Metric	Value
success@reach (first step within threshold — receding-horizon reach controller)	84% (21/25)
success@hold (strict: is_success at the end of the fixed 80-step horizon)	0% (0/25)
mean closest approach gmin	0.020 (threshold = 0.025)

Honest caveat

The policy reliably reaches the goal (84%) but does not hold at it under the fixed-horizon strict metric: after reaching (~step 20) the arm drifts back toward the start, because the training episodes were trimmed to reach+3 steps and provide little "hold-at-goal" signal (and eval zeros the 4 latent state dims, an out-of-distribution condition at the goal). For a deployment that hands off to the next subtask on reach, success@reach is the operative metric.

A hold-tail retrain (E10b) targeting the strict success@hold is in progress: rebuild with a zero-action hold tail (zero-action fraction ~0.30 "sweet spot", vs 0.13 here and 0.67 in the undershooting predecessor E7), plus a latent-dropping ablation.

Recipe headline (the trim fix)

• Predecessor E7 (60-step episodes, 67% zero-action padding tail) undershot: policy magnitude 57% of oracle, gmin0.099, 0% reach.
• Cropping each episode to [0 : t_last+3] (zero-action fraction 0.673 → 0.134) removed the diluted flow-matching marginal (active-phase action mean is 2.66x larger) and cured the undershoot → 84% reach.
• Remaining gap is purely holding at the goal → addressed by the hold-tail sweet-spot retrain.

Not a medical device. Research artifact only.