Quanting Xie's S1 glove records touch and video together
A data collection device aimed at the question of what a million hours of human demonstrations should actually contain.
Quanting Xie has released S1, a skill capture glove for robot learning. It was built alongside a robot hand, and it records human joint movements that map to robot targets, with touch and video on the same timeline.
That last part is the pitch. Most human demonstration data is video of a person doing a thing. S1 logs what the fingers did, what the hand felt, and what the camera saw, all time aligned, so a downstream model gets contact information instead of inferring it from pixels.
The framing is a jab at data scaling
Xie opened the post with a question rather than a spec sheet. "A million hours of what?" he asked, then followed with "If we're scaling human demonstrations for robot learning, what should each hour contain?" The field has spent the last stretch competing on hours collected. He is arguing the unit of measurement is wrong.
If we're scaling human demonstrations for robot learning, what should each hour contain?
The glove being designed next to a specific robot hand matters here. Human hand data is notoriously awkward to use because human fingers do not match robot fingers, so the recorded motion has to be retargeted and some of it is simply not reproducible. Xie says the joint movements map to robot targets, which is the design choosing a narrower, more usable signal over raw human generality.
What has actually been shown
A device, not a result. There are no robot success rates in the post, no policy trained on S1 data, no task demos, no dataset size, no price and no availability details. Nothing in the sources says any robot has yet learned anything from a glove recording. Treat this as hardware for collecting data, and the claim about what that data is worth remains untested in public.
The response that landed hardest came from Humanoid Scott, who replied "It's not the quantity but quality of the data that counts." That is the same argument Xie is making, which is the point. Enough people who work on this now suspect the scaling story for robot data has a quality problem underneath it, and a glove that captures touch is one concrete bet on what a better hour of data looks like.
It's not the quantity but quality of the data that counts.
The open question is whether touch plus aligned video actually trains better policies than a much larger pile of plain video. Nobody has shown that either way yet. Until somebody trains on S1 data and reports numbers against a video-only baseline, this is a well-argued hypothesis with hardware attached.
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.