I was trying to unscrew a bottle cap when my fingers caught on the metal frame. Again.
I was wearing a rigid exoskeleton to collect training data for robots. Picking things up was easy. But as soon as I tried to reposition the cap between my fingertips or twist it open, the hardware got in the way.
The answer seemed obvious: make it soft.
I’ve since come to believe the opposite.
Why wear a glove at all?
You can watch someone tie a tie a hundred times and still fumble your first attempt. You see where the fingers go. You don’t feel how hard they pinch, or which fold of fabric they catch.
Robots have a version of this problem. First-person, or egocentric, video is cheap and abundant. It captures how the world moves and responds, making it a natural fit for training world models.
We’re also getting better at extracting actions from it. EgoEngine turns first-person human video into robot demonstrations, generating both what the robot would see and the actions it could execute. Policies trained on that data transfer to real robots zero-shot, with no real-robot demonstrations used in training. On several tasks, performance approaches that of policies trained on teleoperation data. I expect this line of research to keep improving.
But I believe it has a fundamental limit: fine manipulation that demands millimeter precision.
The precision of an action inferred from video depends on what the camera can see. Contact happens between the fingertip and the object, often hidden by the hand itself. Force isn’t measured in the pixels. A task like tying shoelaces requires precise alignment, handling deformable material, and constant feedback from touch. Better reconstruction can only go so far when that information was never recorded.
Knowing where the fingers are isn’t enough. We need to know what they’re touching, and how.
To learn these skills, robots need high-fidelity action data and tactile feedback.
Figure 1. The camera can see the hand without seeing the contact that matters.
So why not collect the data on robots themselves?
Because teleoperation doesn’t give us a scalable data flywheel. Every operator needs a robot. Collection is limited by what the hardware can do, and more data means more operator time. Costs rise with volume. We need a way to collect more data at a lower marginal cost.
That brings us back to humans. How do we capture what people do with enough precision to train robots?
Put sensors on their hands. Give them data gloves.
Where soft gloves fall short
Conventional data gloves are soft, and that creates three problems.
First, the calibration won’t stay put. Fabric stretches and shifts. Each time you put the glove on, the sensors can sit a few millimeters differently relative to your joints. Your measurements inherit a systematic error that changes from one session to the next.
Second, there’s no fixed geometry for locating contact. Which part of your fingertip is gripping the bottle cap? On a soft glove, the surfaces touching the object are fabric and skin, both of which deform. Better joint sensors won’t tell you where those surfaces have moved.
Third, the tactile signal is hard to get right. Mounting tactile sensors on flexible fabric introduces noise and inconsistency. Getting reliable training data out of that setup is a hard engineering problem.
All three problems stem from the same choice: making the glove soft. We can keep working around that choice, or reconsider it.
What changed my mind
Then I came across ART-Glove, a project from Carnegie Mellon.
It uses sixteen rigid surfaces across the fingers, thumb, and palm. Their geometry is built into the hardware. What caught my attention, though, was how they connect: twenty-two joints aligned with the anatomy of the hand, allowing the surfaces to move with it.
I’d been treating rigidity and dexterity as opposites. The surfaces need to be rigid. The joints need to move.
With rigid links, we can calculate each surface’s position from the joint measurements using forward kinematics. Tactile sensors locate contact on that surface. We no longer have to reconstruct a piece of fabric’s changing shape to know where contact happened. And rigid surfaces give tactile sensors a stable mounting platform, a mature and reliable setup that produces cleaner data.
Making the glove rigid addresses all three problems at their source.
Figure 2. Known geometry and tactile sensing let us locate contact in space.
Back to the bottle cap
That still leaves the problem I started with: the hardware gets in the way.
Adding joints reduces the restriction; it doesn’t eliminate it. Rigid hardware will always impose some limits on the hand, especially during complex in-hand manipulation. That’s the tax you pay.
What changed is that I now think it’s worth paying.
A soft glove gives you more freedom of movement, but sacrifices precision in the measurements you need: motion, contact location, and touch. A rigid glove limits which actions you can capture, but lets you capture them reliably.
You lose some breadth in the data you can collect. You gain confidence in the data you keep. With a soft glove, the underlying measurement problems remain.
I’d rather pay the tax.
The design problem that matters
Once you commit to a rigid glove, the real design work begins. You have to balance four things:
Coverage. Leave too much skin exposed and you’ll miss contacts outside the glove’s known geometry.
Degrees of freedom. The hand needs room to move, with as little loss of dexterity as possible.
Fit. A glove that only works for a narrow range of hand sizes won’t scale.
Comfort. If people can’t wear it for hours, you won’t collect much data.
Improving any one of these puts pressure on the others. The best rigid glove will be the one that gets this balance right.
Figure 3. Coverage, freedom, fit, and comfort have to be designed together.
Egocentric video has taken robot learning a long way, and there’s more to come. But fine manipulation needs a level of contact precision that video alone can’t supply. The next generation of training data needs to combine first-person vision with precise actions and reliable touch.
Data gloves should be rigid. The race is to get those four tradeoffs right.