Dongkkka/Task_000100_lerobot_sim2real_Intern

Task 000100 right MimicGen success 1000, LUT, LeRobot v3

This dataset contains 1,000 successful simulated Task 000100 trajectories selected in episode order from the corrected MimicGen HDF5 source. The source file contains 1,021 successful episodes despite success3000 in its filename; this dataset uses demo_0 through demo_999.

The task is to pick up the navy cosmetic block and place it in the white basket. The LeRobot v3 dataset has 260,859 frames at 15 fps, 22 state values and 22 action values: the original 19 joint values followed by linear_x, linear_y, and angular_z, all set to zero, and three RGB camera videos: cam_left_head, cam_left_wrist, and cam_right_wrist.

The three mobile state and action values are synthetic constant zeros, not measured base velocities or recorded mobile commands. The source HDF5 has no RGB images. Cameras were rendered from recorded pre-step scene and robot state in the Task 000100 WhiteRoom simulation, then transformed with the pinned camera-specific whiteroom_darker_15pct_lut17_v1 LUT before H.264 encoding. This visual replay does not step physics. Source, LUT hashes, and replay details are recorded in meta/isaac_action_replay_provenance.json.

The MP4 shards use 15 fps presentation timestamps. The H.264 packet payloads were preserved during timestamp normalization; see meta/task000100_video_remux.json. Global episode offsets and frame timestamps were checked and repaired where needed; see meta/task000100_episode_offset_repair.json, meta/task000100_video_pts_repair.json, and meta/task000100_validation.json.

The source success flags describe simulated rollouts. Real robot performance has not been evaluated with this dataset.