GelSight Tactile Datasets

High-resolution vision-based touch sensing data from MIT CSAIL. See what robots feel.

Research Only Images + Force Data Tactile Sensing

Overview

GelSight is a family of vision-based tactile sensors developed at MIT CSAIL that use a camera behind a soft elastomer membrane to produce high-resolution images of surface contact. When the sensor presses against an object, the membrane deforms and the camera captures detailed geometry of the contact patch -- including surface normals, texture, and deformation patterns.

Several research datasets have been released using GelSight sensors, covering applications including:

  • Grasp stability prediction: Predicting whether a grasp will succeed from initial tactile contact
  • Texture and material classification: Identifying materials by touch
  • Slip detection: Detecting incipient slip during manipulation
  • Contact geometry reconstruction: Recovering 3D surface geometry from tactile images
  • Cable manipulation: Tactile-guided deformable object manipulation

Access

Note: Most GelSight datasets are released for academic research only. Contact MIT CSAIL for commercial licensing.

Visit GelSight Project

Related datasets

  • Touch and Go -- paired vision-tactile cross-modal data
  • RH20T -- multi-modal manipulation including 200Hz tactile
  • DROID -- visual manipulation at scale

Custom tactile data collection

We instrument robots with GelSight and force-torque sensors for custom tactile data collection.