Datasets:
metadata
language:
- en
license: apache-2.0
tags:
- robotics
- dataset
- lerobot
- ur5e
task_categories:
- robotics
pretty_name: >-
UR5e Dual RealSense Haptic Teleop Dataset (action = q[t+1] absolute, LeRobot
v2.1)
configs:
- config_name: default
data_files:
- data/*/*.parquet
UR5e Dual RealSense Haptic Teleoperation Dataset (LeRobot v2.1)
- Control loop: ~500 Hz (RTDE)
- Dataset sampling: fixed 30 Hz
Definitions:
observation.state[t] = q_actual[t](UR joint positions, rad)action[t] = q_actual[t+1](next-step absolute joint positions at 30 Hz)
Temporal alignment:
- Camera frames are stored in timestamped ring buffers (depth=8).
- At each 30 Hz sample, the frame whose capture timestamp is closest to the RTDE packet timestamp is selected for both cameras independently.
- This eliminates grab-latest race conditions and cross-camera temporal skew.
Visualizer-friendly naming:
- observation.state names: ['shoulder_pan.pos', 'shoulder_lift.pos', 'elbow_flex.pos', 'wrist_1.pos', 'wrist_2.pos', 'wrist_3.pos']
- action names: ['shoulder_pan.next_pos', 'shoulder_lift.next_pos', 'elbow_flex.next_pos', 'wrist_1.next_pos', 'wrist_2.next_pos', 'wrist_3.next_pos']
Videos:
videos/chunk-000/observation.images.wrist/episode_XXXXXX.mp4videos/chunk-000/observation.images.front/episode_XXXXXX.mp4
Per-episode Parquet schema (STRICT order):
action, observation.state, timestamp, frame_index, episode_index, index, task_index