Two New Papers Show Tactile Sensing Is Finally Getting Practical

The second paper, VILAS, takes a different angle. They're building a low-cost modular platform using a Fairino FR5 arm and a Jodell RG52-50 gripper (neither of which I've personally worked with, so I can't speak to their reliability). The clever bit is a kirigami-based soft gripper extension that deforms predictably under load. No explicit force sensing required, just mechanical compliance designed into the structure.

They tested it on grape grasping, which is, look, it's a good benchmark but it's also a bit of a softball. Grapes are fragile but they're also fairly uniform. The real test would be mixed produce or items with variable compliance.

Why hasn't this worked before?

I called my old colleague Heinrich at Siemens last week about something unrelated, and we ended up talking about tactile sensing for twenty minutes. His take, which matches my experience, is that the technology has been "almost ready" since roughly 2012. The sensors work. The control algorithms work. What doesn't work is making it all robust enough for a three-shift operation where the maintenance tech's main qualification is that he showed up.

The TactileReflex approach of deriving thresholds from noise statistics is interesting precisely because it might reduce that maintenance burden. If you don't need to recalibrate every time someone swaps a gripper finger, you've solved half the deployment problem.

The VILAS team is explicitly positioning their system as low-cost and modular, which suggests they're thinking about accessibility. They tested three different vision-language-action models (pi_0, pi_0.5, and GR00T N1.6) on the same hardware, all fine-tuned from publicly available checkpoints. That's the kind of flexibility that makes sense for smaller integrators who can't afford custom development.

What's still missing?

Neither paper addresses what I'd consider the elephant in the room: sensor durability. Vision-based tactile sensors (which both systems use) have come a long way, but they're still more fragile than the industrial force-torque sensors I grew up with. The TactileReflex paper mentions running at roughly 12 Hz, which is adequate for careful manipulation but not for high-speed picking.

There's also the question of what happens when things go wrong. The TactileReflex authors describe their system as a "plug-and-play safety layer" that can sit beneath higher-level control policies. That's a nice idea, but in my experience, safety layers that aren't designed into the system from the ground up tend to create more problems than they solve. You end up with the safety system fighting the main controller, or worse, with operators learning to disable it when it gets in the way.

I should note that I only found these two papers this week, and there may be related work I'm missing. The field moves fast and I don't keep up with arXiv the way I used to.

Is this actually going to ship?

It's too early to say whether either of these specific systems will make it into production. Academic results don't always translate, and both papers are focused on controlled lab environments.

But here's what I find encouraging: the problems they're solving are the right problems. Not "can we make a robot touch things" but "can we make a robot touch things without expensive calibration, custom hardware, or constant babysitting." That's been the gap between lab demos and warehouse floors for as long as I've been in this industry.

If I were still running an integration team, I'd be reaching out to both groups to see what their commercialization timeline looks like. The TactileReflex noise-statistics approach in particular seems like something that could retrofit onto existing systems.

We'll see. I've been wrong before about what's "finally ready." But this feels different than the tactile sensing hype cycles of 2015 or 2019. The underlying VLA models are better, the sensors are cheaper, and the engineering focus has shifted from "impressive demo" to "actually works in the field."

That's progress. Slow progress, but progress.

James Chen · 25 Jun · 5 min