Two Papers That Actually Get Why Industrial Robot Programming Is Still a Nightmare

The second paper, VASO, goes after something harder. It formally verifies that AI-generated robot skills actually satisfy safety requirements before they run. Not "we tested it five times and it worked," but mathematical proof that the skill won't violate temporal safety contracts.

Does any of this actually work?

I'll be honest, I'm skeptical of any robotics paper that doesn't show factory floor results. But the numbers here are interesting enough that I called my old colleague Werner at Siemens to get his take.

PerceptTwin claims about 39% improvement in plan success rates for GPT-5 based planners. That's... actually significant. The paper also claims it improves human verification of failed plans by up to 18%, which matters because someone still has to review these things. The key insight is using an LLM as a "judge" to verify plan correctness, borrowing from the AI alignment literature. Whether that scales to a real production environment remains unclear.

VASO is more ambitious. They're claiming 97.2% compliance with formal safety specifications using fewer than 100 optimization samples. They tested on a Clearpath Jackal (ground robot) and a PX4 quadcopter. Now, neither of those is an industrial manipulator, which is where I'd want to see this applied. But the approach, using formal verification counterexamples as feedback to improve reusable skill contracts, that's genuinely novel.

Werner's take was basically "interesting but we're years away from deploying this." Which sounds about right.

Why should industrial users care?

The dirty secret of industrial robotics is that programming costs dominate. The robot itself might be 30% of a cell's total cost. Programming, integration, and commissioning eat the rest. If you could trust AI-generated programs enough to reduce that programming burden by even 20%, you'd transform the economics of automation for small batch production.

Right now, we don't have that trust. And we shouldn't. I've seen what happens when people deploy untested code (nothing good).

But these papers represent something I haven't seen much of in the LLM-for-robotics hype cycle: actual engineering rigor around the trust problem. PerceptTwin acknowledges that simulations need to match reality, not CAD fantasies. VASO acknowledges that "it worked in testing" isn't good enough for safety-critical systems.

What's still missing?

Neither paper addresses what I'd call the "Tuesday afternoon problem." Your robot program works perfectly for six months, then a new operator loads a slightly different pallet configuration and suddenly your carefully verified plan is garbage. Real industrial environments drift. Fixtures wear. Parts vary. Human coworkers do unexpected things.

VASO's formal verification is only as good as the specifications you write, and writing good specifications is its own nightmare. I spent two months once trying to formally specify "don't hit the operator" in a way that didn't also prevent the robot from doing its job. It's harder than it sounds.

And PerceptTwin's perception-based simulation is limited by perception quality. If your 3D sensors can't see a cable tray that's going to snag your gripper, your simulation won't either.

The bottom line

These are research papers, not products. We're probably three to five years from seeing any of this in commercial deployment, if it works out at all. But I'm more optimistic than I've been in a while.

The LLM hype in robotics has mostly been, frankly, embarrassing. Demos of robots folding laundry that took 47 takes to get right. "General purpose" humanoids that can barely walk. This is different. This is people actually thinking about why we can't trust AI-generated robot programs and trying to build systems that earn that trust.

Will it work? We don't know yet. But at least someone's asking the right questions.

Robert "Bob" Macintosh · 13 hours ago · 3 min