New Research Tackles the Messy Reality of Robot Contact. Here's Why That Matters.

Does the second approach actually work better?

The other paper, Mollified Value Learning, takes a different angle. Instead of trying to enforce strict mathematical constraints at every single point (which apparently gets unstable in complex environments), they're averaging those constraints over a local region. They call it "mollification," which sounds fancy but is basically smoothing out the math so it doesn't explode.

I'll be honest, the theoretical details are above my pay grade at this point. But what caught my attention is that they tested it on "high-dimensional robotic manipulation tasks" and claim it improves goal-reaching performance. The code is open source, which is good. Means people can actually verify the claims.

What I can't tell from either paper is how this performs on actual hardware. Both seem to be primarily simulation-based, which, well, we've all seen how that goes. I called my old colleague at Siemens last week about something unrelated, and he made the same point he's been making for twenty years: simulation is necessary but nowhere near sufficient.

Why should anyone outside academia care?

Because this is the bottleneck. It really is. We've got robots that can see better than ever (thanks to all the computer vision progress), robots that can plan paths efficiently, robots with increasingly capable hardware. But ask them to do something that requires reliable contact, like inserting a cable into a socket or handling produce without bruising it, and you're back to the drawing board.

Warehouse automation is the obvious application. Amazon's been throwing money at this problem for years. The Sparrow system they announced, what, 2022 or 2023, was specifically about picking individual items from bins. Last I heard, it handles maybe 65% of their inventory. That remaining 35% is the contact problem in disguise.

Manufacturing assembly is another one. When I started at Kuka, the running joke was that robots would replace assembly line workers "within five years." That was 1998. We're still waiting, and it's not because the robots aren't fast enough.

So is this actually progress or just more papers?

Honestly? Too early to say. The research direction seems right to me. Acknowledging that contact dynamics are fundamentally different from free-space dynamics, and that you need different learning approaches for each, that matches what I saw on factory floors for twelve years.

But I've seen a lot of promising papers over the decades that never made it out of the lab. The gap between "works in simulation" and "works at 3am when the warehouse is humid and the cardboard is slightly damp" is wider than most researchers appreciate.

What I'd want to see is someone taking this work and testing it on a real Fanuc or ABB arm, with real objects, under real conditions. Until then, it's interesting research with potential. Emphasis on potential.

Mark Kowalski · 9 hours ago · 6 min