Ablation_Cup_visual_raw

Cup-handover visual vs physical DR ablation (data_collection/cup_basket/ablation_vp, 2026-11-07), VISUAL set: raw collector output of every part (24 envs per part, failures included; the LeRobot set Ablation_Cup_visual keeps the first 2,000 successes in part order). Visual randomization only: table look (nominal / flat colour / library texture), backdrop (real lab backgrounds / textures / grid), cup colour jitter, dr2 room lights, 0-3 spot lights over the whole work area. Start pose, table height (29.25 in) and cameras nominal.

Raw Isaac Lab / Isaac Sim 5.1 collector output for the ROBOTIS FFW SG2 (cyclo_lab real2sim pipeline, padiredd). Kept because it has what the converted LeRobot sets do not: full-resolution video, per-episode randomisation / object metadata and the failed attempts (success: false).

folders	113 (abl_cup_visual_p001 ... abl_cup_visual_p113)
episodes	2,712 (2,029 successful)
episodes with video	2,712
files / size	13,787 / 71.5 GB

Converted LeRobot sets (successes only, cameras stretched to 240x240):

• padiredd/Ablation_Cup_visual

Layout

• <folder>/index.json: collector config of that part and its episode list
• <folder>/episode_NNNN/: cam_left_head.mp4, cam_left_wrist.mp4, cam_right_wrist.mp4, meta.json, obs.npz
  • obs.npz (example episode, T steps at 15 fps): timestamp (426,), state (426, 22), action (426, 16), obj_pose (426, 7), tray_pose (426, 7), tcp_pose (426, 7), tcp_target (426, 7), phase (426,), tcp_pose_r (426, 7), tcp_target_r (426, 7), phase_r (426,), arm_clearance (426,)
  • meta.json keys: episode, success, status, cup_init_xy_yaw, tray_init_xy_yaw, cup_end_in_tray_frame, cup_tilt_deg, cup_bottom_z, cup_speed, gripper_l_end, gripper_r_end, final_tags, table_top_z, program_sha, program_result, params, skills, dr, arm_clearance, start_pose, gripper_rand, pair_key, length, fps
• logs/: collector logs
• Cameras: cam_left_head.mp4 (672x376), cam_left_wrist.mp4, cam_right_wrist.mp4 (240x424), H.264.

Uploaded 2026-11-07 14:34 UTC with data_collection/hf_datasets/upload_raw.py.