handvla

Task: grasping. Synthetic DG5FS pregrasp demonstrations, 30 FPS.

Images: target (fixed annotated 224x224 image per episode), RGB, colorized noisy depth. State (25): current absolute joints (rad) 0:20, active thumb/index/middle/ring/little flags 20:25. Action (27): next absolute joints 0:20, world delta xyz (m) 20:23, world relative rotation vector (rad) 23:26, end trigger 26. Rotation: R_next = Exp(delta_rotvec) @ R_current. Terminal frame holds joints, has zero wrist delta, and end_trigger=1. Depth image is colorized, not metric depth. Task ends at pregrasp, not a verified successful grasp. Source and augmentation grouping: source_episodes.jsonl. Split train/validation by original source, not augmentation. This repository is incrementally uploaded; handvla_upload.json records committed progress.