Dongkkka/Task_000100_lerobot_v2_sim2real_Intern

Task 000100 right: 450 episodes across a 5 × 6 grid

This local LeRobot v3 dataset contains 450 selected successful simulation episodes (104,491 frames at 15 Hz). It combines four source sets: MimicGen v2 (112), v3 (112), v4 (113), and the earlier 1000-episode LeRobot set (113). This package has not been uploaded.

All four sources use the same 22 joint/state names and order: 19 robot joints followed by linear_x, linear_y, and angular_z. Every action has head_joint1 = float32(0.7854) and a zero mobile tail. The state mobile tail is also zero. The measured head state can differ slightly from the commanded head angle. All three camera streams use the same whiteroom_darker_15pct_lut17_v1 visual transform.

The grid bins the initial cosmetic position. Rows are X bins from low to high; columns are Y bins from low to high. Each cell first received up to 15 available episodes. The remaining slots were allocated to the least populated cells while keeping the four source counts balanced and maximizing distinct initial randomized scenes.

X bin / Y bin	1	2	3	4	5	6
1	11	17	13	7	1	2
2	16	18	18	18	18	0
3	18	18	18	18	18	6
4	18	18	18	18	18	18
5	18	18	18	18	18	17

Cell X2/Y6 had no successful candidate among the 1,600 source episodes. There are 442 distinct initial scenes among the selected 450 episodes. The grid edges, all episode-to-source mappings, and the selection method are in meta/task000100_grid_selection.json.

The earlier 1000-episode source is task_000100_right_mimic_corrected_success1000_lut_v30. Its source HDF5 trajectory groups are in task_000100_right_mimic_corrected_success3000.hdf5; the first 1000 lexicographically sorted HDF5 group names match the 1000 LeRobot episodes by frame count and sampled joint values. The selected 113 episodes retain the already converted 22D LeRobot actions and states. meta/task000100_v1000_extraction.json records their lossless H.264 clip extraction.

meta/task000100_grid_audit.json records full-frame action/state comparison against staging, grid and source counts, head/mobile checks, and video packet counts. meta/task000100_video_remux.json records unchanged H.264 packet hashes after timestamp normalization; meta/task000100_video_pts_repair.json records synchronization checks. The combined selected HDF5 is alongside this dataset as task_000100_right_mimic_v2v3v4_1000_grid450_joint22.hdf5.

“Successful” refers to the source simulation outcome. It does not claim success after training or on a real robot.