README.md

metadata

license: apache-2.0
task_categories:
  - robotics
tags:
  - lerobot
  - smolvla
  - manipulation
  - handover
  - human-robot-interaction
  - ur3e
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*/*.parquet

SmolVLA Handover Multitasks Dataset

This dataset contains multimodal demonstration trajectories for human-robot interaction and object handover tasks using a UR3e robot equipped with a 2-finger gripper in PyBullet simulation. It is formatted in LeRobotDataset v3.0 format, designed for training Vision-Language-Action (VLA) models such as SmolVLA / OpenVLA.

Dataset Summary

• Total Episodes: 198
• Total Frames: 15,102
• Control Rate: 20 FPS (0.05s per timestep)
• Robot: Universal Robots UR3e with 2-finger gripper
• Observations:
  • observation.images.camera1: High-resolution external angled overview camera (224x224, AV1 video)
  • observation.images.camera2: Top-down workspace overview camera (224x224, AV1 video)
  • observation.images.camera3: End-effector wrist-mounted camera (224x224, AV1 video)
  • observation.state: 6-DoF robot joint positions + gripper state
• Actions: 7-DoF delta end-effector pose [dx, dy, dz, droll, dpitch, dyaw, gripper]
• Tasks & Instructions:
  • handover_direct: Direct object handover to the human receiving hand
  • handover_cube_to_human: Full table pick and handover with dynamic human reaching gestures
  • lift_cube: Object pick-and-lift task