vincent-won/spine-il-hand90

Spine IL — tracked hand demonstrations of endoscopic laminectomy

Imitation learning (IL) for endoscopic spine surgery. One spine surgeon performed 90 left-sided unilateral biportal endoscopic (UBE) laminectomies at L3, L4 and L5 (the 3rd–5th lumbar vertebrae), 30 per level, with a tracked replica endoscope and a tracked replica drill over a hollow skin phantom. Everything under the skin is simulated: the tracked tool poses drive a virtual lumbar spine (L3 to sacrum) that renders the endoscope view and removes bone where the drill goes. No patient data was used.

These demonstrations fine-tuned NVIDIA GR00T N1.7 (vincent-won/spine-il-gr00t-n1.7); two SO-101 arms then executed the policy in closed loop in the same simulation. Code: github.com/geebab/spine-il.

Research and evaluation use only. Not a medical device, not for clinical use or use on patients.

Contents

folder	what
lerobot/	LeRobot v3.0 dataset used for training: 417 episodes, 178,068 frames at 15 fps, 131 instruction sentences
recordings/	the 90 raw hand recordings (recorder schema v2 JSONL, one file per demonstration, by date)
meshes/	level meshes of the virtual anatomy: L3.stl, L4.stl, L5.stl, sacrum.stl (scene frame, mm)
calibration/	robot rig calibration: as-built arm models, hand-eye, marker constellations, tool-tip calibrations (robot and hand tools), camera up vector
robot_runs/	two recorded robot runs (L4, run 17 told to go contralateral, run 19 told to stay ipsilateral) with evaluation caches, run 17's live log and 8 policy frames, for the reproduction scripts in the GitHub repo
recordings_manifest.json	per-recording row counts and SHA-256 of the released files

Demonstrations

• Operator: one surgeon (surgeon1), left side, L3 · L4 · L5 × 30, recorded on three days (16 · 6 · 68 recordings).
• Tracking: Intel RealSense D435 infrared, 848 × 480 at 90 fps; each instrument carries four retro-reflective spheres; pose noise 0.03–0.09 mm. The phantom is registered by touching five skin dimples with the endoscope tip.
• Simulation: bone as a 0.8 mm voxel signed-distance field, spherical burr of radius 2 mm, endoscope view rendered in water.
• Drilling phase: 33.9–218.4 s per recording (mean 103.1 s); removed bone 430.1–2,404.9 mm³ (mean 1,402.9 mm³).
• Each recording keeps the point where each tool's shaft crosses the skin, as the simulator computed it (left working area only).

Labels and instructions

A recording is labelled by how far bone removal crossed the midline (midline minus the 5th percentile of the cut centres' lateral coordinate): ≤ 0 mm → ipsilateral, > 5 mm → contralateral, in between → unspecified. Counts: ipsilateral 31 · contralateral 34 · unspecified 25. Each drilling frame carries one instruction sentence, drawn per frame from 15 wordings per extent (10 for unspecified); 15 % of drilling frames carry a sentence that does not state the extent. Frames around the decision point are duplicated as extra episodes (this is why there are 417 episodes for 90 recordings).

LeRobot features

key	shape	meaning
observation.images.endo	256 × 256 × 3 video	rendered endoscope view (water optics)
observation.state	16	endoscope tip xyz (mm) + quaternion wxyz, drill tip xyz + quaternion wxyz, drill trigger (0/1), pad.zero (always 0); scene frame (LPS, mm)
action	16	next frame's 15 state values + drill.progress (bone removed so far / the demonstration's final removed volume, 0 → 1)
phase.index, cut.progress, track.valid, time.src_s		phase (1–6), cut progress, tracking validity, source time
robot.joints*	12	present in the schema, not used for the hand demonstrations

Robot results (from the paper)

105 closed-loop runs, 35 per level from four start poses. Share of removed bone past the midline: across the three levels, the 48 runs told to go contralateral 9.9 % against 0.4 % for the 48 told to stay ipsilateral (p < 0.001); with wordings never seen in training 9.8 % vs 0.2 %. Per level, unseen wordings: L4 18.1 % vs 0.6 %, L5 10.6 % vs 0.0 % (p < 0.001 each), L3 0.6 % vs 0.1 % (p = 0.053). Contralateral reach was at most 5.2 mm against a demonstrated mean of 10–12 mm, and a supervisory timeout ended 14 of the 48 contralateral runs.

robot_runs/ reproduces the two runs shown in the video: run 17 (the furthest contralateral run) 1069.6 mm³ removed, 26.8 % past the midline; run 19 944.1 mm³, 0.5 %. See robot/eval/judge/ in the GitHub repository.

Not included

• the skin model of the phantom, the viewer / renderer (closed source; offered as a hosted render API during the hackathon judging period), and any patient data (none was used).

Licence

CC BY-NC 4.0. The SO-101 arm meshes in the code repository keep their Apache-2.0 licence (TheRobotStudio).

Citation

@misc{spine_il_2026,
  title  = {Instructed contralateral bone removal in simulated endoscopic spine surgery learned from tracked hand-held demonstrations},
  author = {Won, Donghyun},
  year   = {2026},
  note   = {Dataset, Nebius x NVIDIA Global AI Hackathon 2026. Citation to be updated on publication.}
}