yogyamehrotra/clearvla-sim

ClearVLA simulated lab dataset

2,700 paired scripted-expert episodes (375,360 frames at 10 Hz) of a Franka Panda doing three lab tasks with a test tube that is either opaque or glass in otherwise identical scenes (same seed → same geometry, actions and masks). Rendered with Blender Cycles (real refraction for glass), two 224×224 cameras (fixed diagonal front, hand-mounted wrist). LeRobot v3 format.

Tasks	grasp (put the tube in the rack), pour (pour to the fill line of a beaker; scripted liquid), insert (seat the tube in a 30 mm holder, starts in-hand)
Episodes	300 clean + 150 recovery (DART, 6 mm execution noise, clean labels) per task × material
Observations	observation.images.front, observation.images.wrist (video), observation.state (7 joints + 2 fingers), observation.tcp (xyz, two rotation columns, gripper)
Action	10-D: xyz delta relative to the TCP at the same frame, 6-D rotation (two columns), gripper [0,1]
Extras	material (0 opaque, 1 glass), seed, frame_is_new (a new render every 4th control step; others repeat the last frame), fill_level
Instructions	8 material-neutral paraphrases per task (task field)

Built for the study ClearVLA: do VLA speed-ups fail on transparent lab objects? (Yogya Mehrotra, 2026): code and report at https://github.com/yogyam/clearvla. Simulator details, the sticky-gripper convention and the scripted liquid model are described there. Export script: data/export_lerobot.py.

Citation: Yogya Mehrotra, "ClearVLA: do VLA speed-ups fail on transparent lab objects?", 2026, https://github.com/yogyam/clearvla.