Two New Papers Tackle Robot Safety, But They're Solving Different Problems

VLESA claims a 41 percentage point improvement in action safety through their goal-conditioned approach. That's a big number, though I'd want to know more about their baseline before getting too excited. They tested on something called the ASIMOV-2.0 benchmark, which I hadn't heard of before this week.

The adversarial paper is more honest about where it stands. They explicitly call it "ongoing work" and describe their contribution as a "problem formulation and proposed solution architecture." Translation: they've got a good idea but haven't proven it works yet. I appreciate that kind of candour.

So what

Look, here's the thing. When I was at Kuka, we spent years on safety systems that were, frankly, much simpler than either of these. Light curtains, pressure mats, speed monitoring. Boring stuff that worked because it was boring. The failure modes were obvious and the fixes were mechanical.

These new approaches are trying to solve harder problems (context-aware safety, edge case discovery) but they're also introducing harder failure modes. What happens when the vision system misinterprets intent? What if the Red Team doesn't think of the scenario that actually kills someone?

I called my old colleague at Pilz last month about a different matter, and we ended up talking about ML-based safety systems. His view was that the certification pathway doesn't exist yet. ISO 13849 and IEC 62443 weren't written with neural networks in mind. That remains unclear, how you'd even certify something like VLESA for actual industrial use.

What happens next

Both papers are academic work, and that's fine. The adversarial framework in particular feels like it could be genuinely useful for simulation-based testing, even if it never touches a real robot. The VLESA approach is more ambitious, more likely to face real-world messiness.

I'm skeptical but not dismissive. The egocentric video angle is interesting because it sidesteps some sensor placement problems we used to struggle with. And the adversarial training idea maps onto how good safety engineers already think, just automated.

But we're years away from seeing this in a production environment, if we ever do. The gap between "works on a benchmark" and "works when Dave from third shift is tired and moving too fast" is, well, it's considerable. Anyone who's spent time on a factory floor knows that gap intimately.

The code for VLESA is apparently available on GitHub. Might be worth a look if you're in the research space. For those of us thinking about what to actually deploy next quarter, though, the light curtains are staying put.

Robert "Bob" Macintosh · 5 hours ago · 4 min