Browse, filter, and preview tactile & vision+touch datasets and their episodes.
Coin-sized six-axis force/torque captures on a UMI-style handheld gripper, tuned for contact-rich in-the-wild grasping.
GelSight-instrumented parallel-jaw grasps labeled with grasp success, the canonical touch grasp-outcome dataset.
Self-supervised visuo-tactile pretraining pairs — aligned DIGIT touch and RGB crops for representation learning.
Touch-and-Radiance-Field scenes registering GelSight touches to a neural scene, for spatially grounded tactile.
In-the-wild GelSight touches paired with egocentric video across everyday materials and surfaces.
Implicit neural objects rendering vision, audio, and touch — a simulated multisensory object library.
Tri-modal touch–language–vision alignments pairing tactile readings with natural-language descriptions.
Large-scale cross-modal visual↔tactile pairs on a robot arm for touch-from-sight and sight-from-touch prediction.
Visual-tactile dexterous manipulation benchmark data with per-fingertip touch on a multi-fingered hand.
DIGIT sliding interactions across YCB objects for tactile localization and slip study.
50 hours of bimanual dexterous play on a Dexmate Vega-1 with per-fingertip tactile in three modalities, time-synced to multi-view RGB and language.
Largest bimanual teleop dataset to date — 3,553 hours across 195 tasks with open low-cost hardware and reference models. Vision-only; a target for touch extension.
Diverse humanoid manipulation — 10.3k trajectories across 260 tasks fusing RGB, depth, LiDAR, and tactile with language annotations and a cloud eval platform.
Showing 13 of 13 datasets