Tactile Manipulation Data: The Modality Simulation Cannot Fake
Tactile manipulation data from real contact: per-finger 16x16 taxel force arrays at 100Hz, synced to stereo video, depth and 21-point hand pose. Adding tactile to vision policies lifts task success 8 to 12 percent on contact-rich assembly, measured on a real robot, not a contact-blind sim.
What's in Our Tactile Dataset
Every episode pairs per-finger force with vision on one clock: 16x16 taxels at 100Hz, stereo depth, dual wrist views and 21-point hand pose, so contact is measured, not guessed.
Per-Finger Force Arrays
16x16 taxels at 100Hz on each finger pad — the contact signal simulation cannot fake.
Slip Detection
Onset of slip captured at rates 30Hz video aliases away.
Grasp Stability
Contact distribution that tells a stable grip from a failing one.
Contact-Phase Segmentation
When contact starts and ends, aligned to the video.
Synced to Vision
Force time-aligned sub-frame to stereo video and hand pose.
Deformables
Compliant and deformable object handling that vision alone misses.
Why Vision Alone Misses Contact
A slip that starts and ends inside one 33ms video frame is invisible to 30Hz vision — which is exactly why the force channel exists. Simulation barely models contact, so it under-measures the benefit, and gel-based tactile sensors wear out and stay lab-bound. Three gaps every vision-only or sim-only pipeline leaves open:
We measure contact, not infer it.
Force plus vision on one synced clock: the contact signal a manipulation policy needs, and the one simulation cannot fake. PaXini PXCap taxel arrays record 16x16 force at 100Hz, time-aligned sub-frame to stereo video, depth and 21-point hand pose.
What We Capture, and How
The sensor landscape, honestly compared, plus the sampling rates and delivery formats that make tactile data usable — with the caveats stated up front.
PaXini PXCap taxel arrays
Piezoresistive 16x16 taxels at 100Hz in a 2mm profile, with 100k-plus cycle durability and no gel to wear out. Durable, deployable force capture at scale — the tradeoff GelSight and DIGIT cannot make.
Honest sensor comparison
GelSight and DIGIT are vision-based gel sensors with very high spatial resolution but the gel wears and they stay lab-bound. If your task hinges on fine texture rather than force distribution, GelSight wins on resolution and we are the wrong supplier.
Tactile at 30 to 60Hz
Contact perception is captured at 30 to 60Hz — fast enough to catch the onset of slip and the contact transitions a 30Hz video feed aliases away.
Force-torque up to 500Hz
Whole-arm control needs higher rates, so we capture wrist force-torque up to 500Hz alongside the per-finger tactile channel.
IMU at 200Hz
High-rate inertial capture at 200Hz preserves the fast motion that lower-rate channels miss, so contact-rich and dynamic tasks are represented accurately.
Sub-frame sync to vision
Force is time-aligned sub-frame, on a shared hardware clock, so tactile lines up with stereo video, depth and 21-point hand pose per frame — no post-hoc timestamp guessing.
LeRobot and MCAP delivery
Every episode ships LeRobot-ready, with ROS 2 / Foxglove MCAP and a sync manifest, so force, depth and video line up out of the box — no conversion project before you can train.
Prove it on a real robot
One honest caveat: tactile needs paired vision to be useful, and sim benchmarks like LIBERO and SIMPLER barely model contact, so they under-measure the benefit. The 8 to 12 percent lift has to be shown on a real robot.
Adding Contact to Your Policy?
Tell us the manipulation task. We'll map the tactile capture and send a representative sample pack — the same taxel arrays, sampling rates and delivery format you'd get in production, so your team can inspect force-to-vision alignment before any commitment.
Book a Strategy Session →Where Our Tactile Data Gets Used
Contact-rich tasks where force closes the loop and vision-only policies fall short — captured to your taxonomy, not a generic corpus:
Send your manipulation task and we scope a tactile capture program against it. Force plus vision on one synced clock is the contact signal a manipulation policy needs, and the one simulation cannot fake — so the data you train on is measured on real contact, not inferred from contact-blind video or sim.
Frequently Asked Questions
GelSight gives higher spatial resolution but the gel wears and it stays lab-bound. PaXini taxel arrays give durable, deployable force at 100Hz with no gel. We use taxel arrays for field capture at scale; if you need fine texture, GelSight is the better fit.
Yes. Adding force and tactile to vision policies reports an 8 to 12 percent success lift on contact-rich assembly. Because sim barely models contact, that lift has to be shown on a real robot.
30 to 60Hz for tactile contact perception, up to 500Hz for whole-arm force-torque control, and 200Hz IMU. We capture all three.
Sub-frame, on a shared hardware clock, with a sync manifest, so force lines up with stereo video, depth and hand pose per frame.
Per-dataset and non-exclusive by default, permitting commercial training, with exclusive and custom-capture terms on request.
Related Solutions in the Physical-AI Cluster
Physical AI Training Data — Multimodal Field Capture for Robot Learning
Learn more →World Model Training Data — Real Video for World Foundation Models
Learn more →Embodied AI Training Data — Demonstrations for Robot Policies & VLAs
Learn more →Robot Data Collection Services — Teleoperation & Egocentric Capture
Learn more →Physical AI & Robotics — Consulting, Development & Training Data
Learn more →Private & On-Premise AI — Self-Hosted AI Deployment
Learn more →AI Transformation Workshop
Half-day strategy workshop to map your contact-rich manipulation tasks and identify the right tactile capture program for your policy stack.
Book a workshop →AI Strategy Session
30-minute scoping call. We'll talk through your manipulation task, target model and delivery format, then scope a representative tactile sample pack.
Book a session →AI Consultant vs In-House Team
Honest tradeoffs on standing up tactile field capture in-house versus engaging a partner for a scoped, delivered capture program.
Read the comparison →See the Data Before You Decide
A 30-minute strategy call. We'll walk through your manipulation task, target model and delivery format — then scope a representative sample pack so your team can inspect force-to-vision alignment and prove the lift on a real robot.