Tactile data for robot manipulation
Robots can see. They still can't feel.
We record human hands performing real tasks. Each session pairs full-hand pressure across 200 taxels with egocentric RGB-D and 3D hand pose. We're working with teams training manipulation policies to cater exactly to their specifications.
Our datasets
- Full-hand pressure
- Normal pressure across 200 taxels covering the fingers and palm, each with 12-bit precision.
- Egocentric RGB-D
- Color and per-pixel depth on the same clock, so a contact event can be trained from with no ambiguity.
- 3D hand pose
- Per-frame 3D hand pose helps fully characterize the context behind each grasp event.
- Task structure
- Post-processing subdivides each task and annotates with action, object, and grasp.
Why touch
Vision defines object positions, but only where it can see. Touch fills in the gaps and defines grip strength.
A policy trained only on pixels has to infer contact from appearance, which is difficult for soft bodies and heavy objects. This inference becomes impossible the moment the hand occludes the thing it is holding.
Manipulation datasets today are still overwhelmingly visual, because touch is harder to instrument. Frontier labs use brute-force to learn touch through expensive, slow teleoperation. Diverse, scalable tactile datasets step in next to vastly improve training efficiency.
From hand to dataset
Each demo is rigorously tested
Capture
Operators wear our hardware and perform tasks. Low-friction equipment guarantees they can work comfortably to provide high-confidence recordings.
Review
Verification pipelines test each session, filtering out low-fidelity or compromised demonstrations.
Release
Approved sessions are aggregated and delivered as MCAP logs, or as a LeRobot dataset that loads into an existing training pipeline.
Train on the signal you're missing.
We're taking on a small number of partner labs for the first release. Early partners help shape which tasks we capture.