Two New Papers Show Robots Learning to Actually Touch Things (Finally)

The second paper attacks a related but distinct problem: how do you make simulators that actually match reality without needing a warehouse full of training data? This is the real-to-sim gap, and if you've spent any time in robotics you know it's where good ideas go to die. Your policy works perfectly in simulation, you transfer it to the real robot, and suddenly everything falls apart because the simulated friction coefficient was off by 0.1.

The team behind this work, also on arXiv, proposes what they call a "few-shot" approach. Instead of collecting thousands of real-world examples (expensive, slow, and often dangerous), they use a small amount of real data to calibrate an analytical simulator. That calibrated simulator then generates synthetic training data. A mesh-based graph neural network learns the forward dynamics, and here's the clever bit, they derive surrogate gradients for collision detection to make the whole thing differentiable.

For the non-technical readers: differentiable means you can use gradient-based optimization, which is basically how modern AI learns anything. Making collision detection differentiable is hard because collisions are discontinuous events, either you're touching or you're not. There's no smooth gradient to follow. Their workaround appears to work, at least in their experiments.

The results claim they outperform "differentiable baselines" in replicating real-world trajectories. I'd want to see more independent validation before getting too excited, but the direction is promising.

So what does this mean for anyone actually building robots? Honestly, probably not much in the immediate term. Both papers are research contributions, not products. The spray painting framework was tested on two object shapes in simulation. The few-shot simulator was validated in "multi-object interaction scenarios," which is academic speak for "we pushed some blocks around."

But I think there's something important happening here that's worth paying attention to. For years, the robotics field has been split between two camps: the classical motion planning folks who can guarantee geometric correctness but produce movements that look, well, robotic, and the learning-from-demonstration crowd who can capture human-like motion but only for the exact scenario they trained on. These papers represent attempts to bridge that gap.

The first paper decouples geometry from execution expertise. The second paper tries to make simulation accurate enough that skills learned in simulation actually transfer. Both are chipping away at the same fundamental problem: robots need to understand physics the way humans do, intuitively and flexibly, not just mathematically.

I've covered three different tech verticals since the 90s, and one pattern I keep seeing is that the unsexy infrastructure problems get solved right before the flashy applications take off. Nobody cared about TCP/IP until suddenly everyone needed the internet. Nobody cared about smartphone batteries until suddenly everyone needed smartphones. And maybe, just maybe, nobody cares about surface-interaction dynamics and real-to-sim transfer until suddenly everyone needs robots that can actually do useful physical work.

Or maybe these are just two more papers that'll get cited a few dozen times and forgotten. It's too early to say! But if you're placing bets on where robotics research needs to go, "making robots understand how touching things actually works" seems like a reasonable horse to back.

The companies deploying industrial robots today are still mostly programming paths by hand or using crude teach pendants. That's roughly 25,000 new industrial robots installed in the US per quarter, according to the latest Association for Advancing Automation numbers, and most of them are doing pick-and-place or welding along predefined trajectories. The kind of adaptive, transferable skills these papers are working toward would change that equation significantly.

But what do I know. I still prefer email to Slack, and half the young founders I talk to think that makes me a dinosaur. If you want to argue about whether this research matters, my email's on the about page.

James Chen · 2 hours ago · 7 min