Contact-rich tasks that actually work.
Pilot teams adding 6thSense tactile streams to existing demonstrations saw dish-loading success rise from 44% to 75% with no model changes.†
Robots that feel what they’re holding.
Demonstrators wear a glove that captures touch. Your robot wears tactile skin molded to its own hand. Every episode ships with pressure, hand pose, depth, and video on one clock — ready to train on.
Trusted by robot-learning teams atHalden Robotics†Corvid Labs†Meridian Dynamics†Osgood Institute†Fieldline AI†
The 6thSense capture glove, palm open.
- 0.0M
- aligned episodes delivered†
- <1 ms
- cross-modal sync
- 0
- pilot teams capturing touch†
Winner · 2026 Manipulation Data AwardOpen Dexterity Consortium†
Four ways your dataset is touch-blind.
- 01
Your policy fails at the moment of contact.
It reaches fine, then fumbles the grip — because the demonstrations never recorded when touch began, how hard it pressed, or how the grasp adjusted.
- 02
Your sensors produce folders, not datasets.
Off-the-shelf tactile pads and a recording script give you raw dumps with drifting clocks, uncalibrated channels, and hours of cleanup per hour of capture.
- 03
You’ve spent a quarter building a rig instead of collecting data.
Sourcing sensors, wiring, debugging sync — none of it is the research.
- 04
Scale isn’t fixing it.
Contact-rich manipulation is data-constrained. Bigger models trained on the same touch-blind demonstrations plateau in the same places.
Touch can’t be bolted on after the fact.
It has to be captured at the source, on the same clock as everything else, with its limits stated.
So we built the glove and the robot skin as one system. Sync and calibration live in the hardware, not in a script you maintain. And every channel comes with a documented boundary: where it’s reliable, how drift is handled, and what should never be mistaken for ground-truth force.
You get touch you can train on — and you know exactly what it means.
Touch · video · depth · poseTimestamps to repairOne clock
How it works.
01
Capture
Demonstrators put on the glove and egocentric rig and do the task. Your robot’s molded skin records the same contact signals during rollouts.
Recording starts on day one, on real household tasks.
02
Sync
Tactile, hand pose, RGB-D, IMU, and secondary cameras are stamped against a single clock at capture time.
No post-hoc alignment, no timestamp archaeology.
03
Calibrate
Per-channel checks for drift, fit, and timing run on every session, with results attached to the episode.
You know which signals to trust before you train.
04
Package
Episodes land in the format your trainer expects, with contact phases, subtask boundaries, outcome labels, and QC flags.
Aligned episodes, not raw folders.
One episode, aligned
Touch, video, depth, and pose on one clock. Pick a task.
What changes when touch is in the dataset.
3 daysunboxing to first episode
First usable data in days, not quarters.
Median time from unboxing to first packaged episode: 3 days. Typical in-house rig build: 3–6 months.†
<2%episodes fail QC
Almost nothing thrown away.
Under 2% of 6thSense episodes fail QC, versus roughly 30% for DIY tactile stacks — so an hour of capture is an hour of training data.†
Ready to give your robot a sense of touch?
Teams already training on touch.
71%from 38%
fold-and-stack success
Trained on 2,100 glove-captured episodes with contact phases. Fold-and-stack success went from 38% to 71%; policy stopped crushing garments on the first grasp.
<1hfrom 11h a day
dataset prep per capture day
Replaced a homegrown tactile rig with the 6thSense stack. Cut dataset prep from 11 hours per capture day to under 1, and stopped losing sessions to clock drift.
“Before, contact timing was a guess we made from video. Now it’s a channel. That one change moved our grasp success more than the last three architecture experiments combined.”
Four fair objections, answered straight.
Our answer
They aren’t, and we say so. But policies don’t need newtons — they need to know when contact started and how it trended. That’s exactly what the channels capture, with the boundaries documented so you never over-trust them.
The practical questions.
LeRobot, RLDS, and HDF5, with a versioned schema. Custom exporters available on enterprise plans.
Robots that feel what they’re holding — starting this month.
Capture touch at the source, on one clock, with its limits stated. Get episodes you can train on from day one.
Two-week pilot. If the first packaged episodes don’t meet the QC bar we agree on up front, you don’t pay.†
Book a capture pilot.
Tell us the tasks you’re training, the hand your robot uses, and what your trainer reads. We’ll agree on the QC bar up front and set a start date.
- Hour 1Glove and ego rig set up
- Half a dayRobot skin mounted and calibrated
- Day 1First episode recorded
- Week 2Pilot ends. No long-term commitment.
- One glove · one fitted skin
- You own the data · we never train on it