Dexterous manipulation data
Demonstration hours, at a fraction of the cost.
Tell us the manipulation task you need a robot to learn, and we record it: expert hands, a $140 instrumented glove, the time series in your hands. Teleoperation charges $30–100 an hour for the same thing.
Working prototype · recording with our first design partners
- 5
- digits tracked
- 60 Hz
- sample rate
- 14
- calibration gestures
- $140
- per glove
The proof
One take, two views, same frame.
Left: the glove on a hand. Right: what the software reconstructed from its channels at that instant, thumb included. Both recordings come from one session, aligned by cross-correlating motion — nothing is re-enacted.
The bottleneck
Teleoperation doesn't scale.
Each usable hour costs an operator's hour, a robot's uptime, and a rig rebuilt per task. A cheap glove breaks that curve: hand it to a thousand people who already have the skill, and they keep doing their own work while it records.
The data
Time series from a hand, not a bounding box.
Rotary potentiometers sit on a 3D-printed exoskeleton at each joint. No cameras, no occlusion, no inference between the motion and the number.
One glove, any hand
The fit is solved in software, not in plastic. Fourteen calibration gestures measure the wearer, and the same hardware moves on to the next person — no resizing, no reprinting, no per-operator rig.
Retargets to your hardware
The retarget is solved per target hand, so a robot thumb built to different geometry is a configuration rather than a rewrite — and the thumb is the joint that usually breaks when human demonstrations meet real hardware. Tell us which hand you are targeting. It runs against a virtual hand today.
What a recorded session contains
- Per-joint flexion angles for the thumb and all four fingers
- Full session time series, sampled continuously, not keyframed
- The calibration solved for that wearer, stored with the session
- Session metadata: who wore it, what the task was, which trial
- Synchronised video of the same session for grounding and review
Limits
What we don't capture yet.
You would find these out in the first ten minutes of an evaluation. If one is disqualifying for your work, tell us — it moves up the roadmap.
Wrist pose in world space
Next revisionPotentiometers measure joint angles, not where the hand is. A 6-DoF wrist tracker is the next hardware revision.
Contact forces
RoadmapWe record how the hand moved, not how hard it pressed. Fingertip force sensing is the next sensor we want to add.
Object and scene state
Partner-dependentSessions ship with synchronised video, but we do not currently provide segmented object poses.
Why not vision
Camera-based hand pose is free and fine for open-air gestures. It degrades exactly where the interesting work happens: fingers inside a material, hands occluded by the tool, sub-centimetre motion in a cramped workspace.
Where this goes
Skills that were never written down.
A joiner knows how much pressure a chisel takes before the grain tears. None of that lives in a video dataset, because none of it is visible — it lives in the hand. Putting gloves on experts while they work is what we are building towards, and the domains below are where we want to start.
- Joinery
- Watchmaking
- Leatherwork
- Kitchen work
- Your domain
Access
Tell us what to record.
Name the task and we come back with what it would take to record it: which expert, how many hours, and what the first batch looks like. Craftsmen willing to be recorded are equally welcome — you are the other half of this.