eunyoung927's picture
Add dataset card
0a79f12 verified
|
Raw
History Blame Contribute Delete
1.7 kB
metadata
license: apache-2.0
task_categories:
  - robotics
tags:
  - LeRobot
  - lerobot
  - lerobot-v3
  - ur7e
  - real-robot
  - robot-manipulation
  - imitation-learning
  - vision-language-action
  - vla
  - cap
  - pick-and-place
  - h264
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*/*.parquet

UR7e CaP Pick-and-Place 100 Episodes

This dataset contains 100 successful forward episodes from a UR7e real-robot CaP pick-and-place session, converted to LeRobot v3 format.

  • Repository: CoRL2026-CSI/UR7e_CaP_PickandPlace_100epi
  • Source session: /home/franka/hscho_workspace/AutoDataCollector-ur7e-v2/results/ur7e/session_20260417_135807
  • Selection rule: first 100 forward episodes with judge_result.prediction == "TRUE"
  • Task prompt: pick up the red block and place it on the blue dish
  • Episodes: 100
  • Frames: 107,523
  • FPS: 30
  • Robot: UR7e
  • Video cameras: observation.images.realsense_topview, observation.images.realsense_wrist
  • Video encoding: H.264, yuv420p, 640x480, 30 FPS
  • Reset episodes: excluded

Schema Notes

observation.state and action are 7D vectors: six UR arm joints in radians plus the gripper position. The action field preserves the logged absolute joint target representation from the source collector.

Additional task/skill context is included in skill.* columns, including skill type, progress, language, joint goal, robot/world pose goals, and gripper goal.

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("CoRL2026-CSI/UR7e_CaP_PickandPlace_100epi")

Provenance

See meta/source_session.json for the selected source episode names and per-episode frame counts.