Tactile Sensing Research Is Finally Getting Serious, and It's About Time

Ten milliseconds! That's fast enough to actually be useful. I've seen this movie before with computer vision, where academic demos ran at one frame per second for years before anyone figured out how to make them real-time. The fact that tactile sensing research is already hitting these speeds suggests the field has learned something from those earlier cycles.

Mixing sensors like a DJ mixes tracks

The second paper takes a different approach, and honestly it's the more ambitious one. Researchers from what appears to be a multi-institution collaboration (the paper lists a project page at mitas-touch.github.io) asked a simple question: what if robots used multiple tactile sensors at once, the way humans use different types of nerve endings?

They call it Multi-Resolution Tactile Sensing, or MiTaS, and it combines three inputs: a regular RGB camera, a vision-based GelSight Mini sensor (which basically photographs the deformation of a gel pad when it touches something), and a high-frequency event-based Evetac sensor. Each of these operates at different temporal resolutions, some capturing slow changes, others catching fast ones.

The results are striking. Vision-only approaches solved the manipulation tasks 31% of the time. Adding one tactile sensor bumped that to 54%. But the full multi-sensor system hit 80% success rates across five different contact-rich tasks.

Call me old-fashioned, but I find these numbers more impressive than most of the headline-grabbing robot demos I've seen lately. A 49 percentage point improvement from adding touch? That's not incremental. That's the kind of gap that suggests we've been doing something fundamentally wrong by ignoring tactile feedback for so long.

So what

I can already hear the objections. These are lab results. The sensors are expensive. The tasks are still relatively simple compared to, say, a warehouse worker sorting packages. All true!

But here's the thing: this is exactly how useful technology develops. You get the fundamentals working in controlled conditions, you figure out what representations and architectures actually work, and then the engineering grind begins to make it cheaper and more robust. The self-driving car hype cycle taught us that skipping the fundamentals leads to, well, robotaxis that still can't handle rain.

What I find encouraging about both papers is their focus on practical constraints. The NeuroTac work explicitly measures latency because the researchers understand that a sensor that takes 100 milliseconds to process is useless for real manipulation. The MiTaS team even tested whether training with multi-sensor data helps when you only have single sensors at deployment time (it does, boosting performance by over 10% in some tasks). These are the kinds of questions that matter if you actually want robots to work outside a lab.

The other thing worth noting, and this is something the young founders chasing humanoid hype might want to consider, is that neither paper requires a humanoid form factor. These sensing approaches could work on any robot that needs to manipulate objects. Industrial arms. Mobile manipulators. Prosthetic hands. The research is about the sensing and learning, not about making robots look like people.

I've been covering robotics long enough to be skeptical of most claims, but what do I know. Maybe the humanoids will figure out touch on their own. Or maybe the field will eventually realize that the fundamental research happening in labs like these is what actually moves the needle. The papers aren't exciting. They won't get venture capital pitches or TechCrunch headlines. But they're solving real problems, and they're solving them in ways that might actually scale.

If you want to argue about any of this, my email's on the about page. I still check it more than Slack.

Mark Kowalski · 3 days ago · 6 min