Hacker News Reader

Comment by eichin | Hacker News Reader

Very nice, and much easier to manufacture than the old Takktile sensors https://biorobotics.harvard.edu/takktile.html - it also looks like you could use the skins to destructive levels of force, without damaging the circuit boards at all, with a stiff enough layer between the chips and the skin (the Takktile system put the epoxy directly in contact with the pressure sensors, so while you could use protective layers over that, it would necessarily reduce the sensitivity.)

How tech-independent is the policy learning part? Do the models end up relying on how the board is giving you direction vectors, rather than contact location? (Nothing wrong with that, I'm just wondering if the directional aspect "factors out" certain kinds of change, and thus simplifies the learning process.)

View on HN

While the sensor gives us direction vectors, they serve as good proxies for contact location, as we showed with ReSkin, https://reskin.dev.

That being said, the exact quantities the policy depends on are hard to interpret, given the use of deep learning. This could potentially be modality agnostic, but there has been no sensor so far that has shown (1) the ability to detect intuitively relevant quantities like contact location and 3-axis forces, and (2) sufficient signal consistency for deep learning models to generalize across instances. This was a key motivating factor for AnySkin, and we found a relatively straightforward fabrication procedure that enables this for magnetic sensing.

Curious, could you not calibrate using a force sensor, then include the output as a learning parameter. This seams a naive approach, which likely means it has been tried early on with other low hanging fruit, but I'm curious what the analysis of that approach is. Is there a fundamental reason this wouldn't work?

You could, and this is what we did with ReSkin, https://ReSkin.dev

The reason we don't want to do this is that it is difficult to cover all possible characteristics. Say we do single point contact localization, and 3-axis forces prediction. What happens when we have multi-point contact? The calibration has only been used to calibrate/align in a lower dimensional space. This is primarily why not needing calibration and baking this into the hardware is a lot more appealing. The user/designer no longer needs to think about the task and the dimensions of alignment required for that task.

Reply on news.ycombinator.com