ajaysri/lego_flat_final_goal_200_v1

LEGO flat final-goal dataset

200 robot demonstrations: 100 human-teleoperated episodes + 100 correction episodes, 79,956 frames at 20 Hz. This is simulated robot data; it is not the synthetic human-hand proxy dataset.

Use the visual_inspection viewer to browse every episode: front camera at start, middle and end, middle wrist camera, final goal image, and its text description. Full front and wrist MP4 recordings are in videos/.

Training representation

The fixed goal image depicts the complete final assembly, using an isometric view and layer cross-sections. The fixed text describes the same complete structure. Both are stored for every frame. Training chooses exactly one per example: 50% image-only, 50% text-only. Image-only uses a generic assembly instruction; text-only zeros and masks the goal-image slot. Current front/wrist images and robot state remain available in both cases. There is no next-subtask conditioning.

The pi0.7 run uses a 20-step action horizon and 256 text tokens. Its owner-requested target was extended to 16,000 optimizer updates, retaining checkpoints every 2,000 updates. The dataset is unchanged.

Data integrity

State, action, and timestamp values were checked against the 200-episode source. Current-camera videos are preserved, with no new overlays or image cleanup. White regions near the image boundary are present in the source renderings. The final-goal image is intentionally a schematic, not a camera frame.

The real LeRobot reader loaded all 200 episodes. Representative camera/goal grids were visually inspected; this is not a manual review of all 79,956 frames. The inspection table is for browsing only and is not additional training data. meta/derivation.json records episode membership and derivation hashes.

All episodes are training data. This release defines no held-out evaluation split.