Does a robot need to be told how hard a human pressed?

Human video now scales. A million hours of egocentric footage are free to download, and published work shows manipulation performance rising with hours of it.

What video does not record is contact. How hard a hand pressed. When contact began. Whether the material slipped.

Two lines of recent work disagree about whether that gap matters. One argues contact has to be measured directly. The other argues it can be estimated from pixels well enough that measuring is wasted effort. Nobody has run the cheap version of that comparison on deformable materials, because nobody has the data to run it on.

The work still done by hand here survived because machines could not do it. Cloth, leather, wire, sheet brass. Deformable, contact rich, two handed, irregular.

That is not a development lag. It is a filter, and it selected for exactly the problems robotics has not solved. I grew up around it. My father runs a brass workshop in Moradabad.

A wearable that records what the hand sees and what the hand feels, on one clock.

A head camera for egocentric video. A forearm band with dry electrodes reading muscle activity, calibrated against measured force. Fingertips stay open, because a tailor who cannot feel fabric takes the rig off in ten minutes and then there is no data at all.

Then a few hundred hours in real workshops, and one experiment: train the same policy twice, once with the force channel and once without, and measure the difference. Release everything openly.

If the force conditioned policy does not beat the video only policy on held out tasks, the thesis is dead and I publish that result. The hypotheses and the thresholds are written down in advance on the research page so I cannot move the goalposts afterwards.

Three other ways this fails, in order of how likely I think they are:

1. Muscle signal is too noisy to recover useful force in a real workshop. Sweat, movement, and the mapping shifting every time the band comes off and goes back on.
2. Force turns out to matter less than shear, which this version does not measure.
3. Workers will not wear it for a full shift, and a rig nobody wears collects nothing.

Nothing collected in a workshop yet. Everything below is on a table in Bengaluru.

Time, and people to argue with. The experiment itself is small and cheap and I can run it alone. The part I cannot do alone is working out whether I am asking the right question.

If you train manipulation policies and there is a material, task or force profile you need, say so and we will collect for it.