LASK: A Dataset for Laparoscopic Skill and 7-DoF Kinematics

Accurate perception of surgical instruments is crucial for automated skill assessment and computer-assisted training in minimally invasive surgery, yet publicly available video-kinematic datasets remain scarce, particularly for non-in-vivo tasks. We introduce LASK (LAparoscopic Skill & Kinematics), a peg-transfer box-trainer dataset that synchronises endoscopic video with dense ground-truth instrument kinematics and manual instrument annotations for two graspers tracked throughout. LASK comprises 37 annotated trials organised into a cross-cohort benchmark: a 7-DoF training set (19 trials), a 7-DoF in-distribution validation set (10 trials), and a 6-DoF out-of-distribution test set (8 trials), spanning surgeons from novice to expert. Each trial provides per-frame electromagnetic tool poses (~91,000 frames; positions in millimetres, orientations as unit quaternions, plus calibrated jaw angles for the 7-DoF cohorts), time-aligned to the video. On top of this, manually labelled keyframes supply instrument segmentation masks, tooltip and jaw (left/right) keypoints, shaft-joint keypoints, and per-component visibility flags for both instruments. The release further includes surgeon-specific metadata — handedness (left/right/both) and procedure experience (lifetime and last 12 months). With two instruments captured continuously under wide-field box-trainer imaging, LASK supports robust benchmarking of multi-class instrument detection, segmentation, tracking, pose estimation, and surgical-skill classification, and its 6-/7-DoF split enables study of cross-configuration generalisation. This is a current subset of the entire dataset, which will be added to this record in the near future.