Episodes Preview LeRobot Visualizer
5 episodes · 15 fps · 5 cameras · 640×360 avc1

RoboLab Tactile Test Dataset

Simulated robot manipulation dataset with GelSight tactile sensor data, collected in NVIDIA IsaacLab.

Overview

Property Value
Task Pick up banana and place in bowl
Robot Franka Panda + Robotiq 2F-85 gripper
Policy π0.5 (pi0.5_droid_jointpos)
Episodes 5
Total frames 715
FPS 15 Hz
Success rate 100% (5/5)
Format LeRobot v3.0

Sensors

  • external_cam: 360×640 RGB (third-person view)
  • wrist_cam: 360×640 RGB (wrist-mounted)
  • ego: 240×432 RGB (egocentric mirrored)
  • tactile_left: 320×240 RGB (GelSight on left finger)
  • tactile_right: 320×240 RGB (GelSight on right finger)

Tactile Setup

GelSight Mini-style sensors simulated on both fingers of the Robotiq 2F-85 gripper:

  • Compliant fingertip contact (stiffness=200, damping=1.0)
  • Backstop depth: 3.0mm
  • Camera focal length: 6.9mm
  • Pad cameras positioned at empirically tuned offsets

Features

Feature Type Shape
action float32 [8]
observation.state float32 [13]
observation.velocity float32 [13]
observation.ee_position float32 [3]
observation.ee_orientation float32 [4]
observation.images.external_cam video [360, 640, 3]
observation.images.wrist_cam video [360, 640, 3]
observation.images.ego video [240, 432, 3]
observation.images.tactile_left video [320, 240, 3]
observation.images.tactile_right video [320, 240, 3]

Collection

Collected using RoboLab with the π0.5 vision-language-action policy serving actions via gRPC. The tactile sensors are integrated into the Robotiq gripper fingers using compliant contact simulation in PhysX.

Downloads last month
48