ur5e-multitask-5obj-275ep-1009

UR5e real-robot teleoperation, 5-object multitask dataset (LeRobot v2.1). Merge of the previous 251-episode multitask set with 24 newly-recorded 2026-10-09 episodes.

• Robot: UR5e + Robotiq 2F-85 gripper (real)
• Episodes: 275 (187,995 frames @ 30 fps, 825 videos)
• State / Action: 7-dim each = 6 arm joints (deg) + 1 gripper (0.0 open → 1.0 close)
• Cameras (640×480, video, real): observation.images.front / side / top

Tasks (per-episode task_index)

idx	task	episodes
0	Pick the Erlenmeyer flask and put it into the tray.	53
1	Pick the graduated cylinder and put it into the tray.	53
2	Pick the transparent vial and put it into the tray.	52
3	Pick the test tube and put it into the tray.	54
4	Pick the beaker and put it into the tray.	63

Composition

• ur5e-multitask-5obj-newroom-add1005 (251 ep, multitask base)
• ur5e-beaker-3cam-adversarial1009 (12 ep, beaker → task_index 4)
• ur5e-transparent-vial-3cam-1009 (12 ep, transparent vial → task_index 2)

Merged with merge_multitask_tasks.py: task_index remapped by task string to the canonical 5-task set (so the single-task batches land on the correct multitask index, not raw index 0); episode_index / global index / video filenames re-sequenced; meta rebuilt.

Modality

GR00T modality.json included: single_arm = state/action[0:6], gripper = [6:7], video keys front/side/top, human.task_description from task_index. meta/stats.json + relative_stats.json generated.

Notes

• All real (USE_REAL_CAMERA=True); front camera re-calibrated 2026-10-09 (reproj mean 0.7 px).
• parquet task_index ↔ episodes.jsonl task strings verified consistent at merge boundaries.
• Intended for the next 5-object multitask GR00T finetune (successor to the 251-ep set used for indojin/ur5e-multitask-5obj-newroom-add1005-120000).