Two new papers suggest expensive sensors might be optional for robot manipulation

They're using what they call a "physics-guided Transformer" that incorporates Newton's second law and Coulomb friction directly into the loss function. The claim is that this reduces estimation error by over 10% compared to methods that actually have access to full force information.

You might be wondering: how does knowing less give you better results? The paper argues it's about generalization. By baking physics constraints into the model, it handles out-of-domain conditions (new objects, new surfaces) more reliably than pure data-driven approaches.

Tbh, I'm not sure I fully buy the 10% improvement claim without seeing more detailed comparisons. The simulation-to-real gap is always tricky, and "challenging out-of-domain conditions" can mean a lot of things.

So what does this mean?

Here's what I think is actually happening: researchers are getting creative because the hardware path to better manipulation is expensive and slow. Force/torque sensors that work reliably in real deployments cost real money. Tactile sensing is still mostly a research curiosity. Multi-camera setups are a nightmare to calibrate and maintain.

So people are asking: what can we do with less?

The VILAS approach says: build compliance into the mechanics. The PhyPush approach says: extract more information from the data you already have.

Neither of these is going to replace proper sensing for every application. If you're doing surgery or handling explosives, you probably want actual force feedback. But for the vast majority of manipulation tasks (picking things up, moving them around, placing them somewhere), maybe good enough is good enough.

What's still unclear

A few things I'd want to know before getting too excited:

The VILAS paper doesn't give exact cost figures for their platform. "Low-cost" means different things to different people. The Fairino FR5 they're using isn't exactly cheap by hobbyist standards.

PhyPush was validated on pushing tasks specifically. It's not obvious how well the approach generalizes to other manipulation primitives. Pushing an object gives you a lot of information about mass and friction. Lifting it? Different story.

And honestly, both papers are still in the "research demo" phase. The gap between "works in the lab" and "works in a warehouse at 2am when something goes wrong" remains enormous.

But the direction feels right. If we can get 80% of the capability at 20% of the hardware cost, that changes what's economically viable. And in robotics, economics eventually wins.

Mark Kowalski · 10 hours ago · 5 min