Two New Papers Remind Us That Robot 'Success' Isn't Always What It Looks Like

When I was at Kuka, we used to joke that the robots had no short-term memory and no shame. They'd repeat the same mistake indefinitely without any sense that they'd already tried that. This research is, in a roundabout way, working on the memory and inference side of that problem. It remains unclear how well the approach scales beyond the relatively constrained task types they tested, and the paper is honest about that.

The second paper, SafeManip, is the one I think deserves more attention in industrial circles. The core argument is simple and a bit uncomfortable: task success doesn't guarantee safe execution. A robot can complete a task and still behave unsafely during that same run. The authors built a benchmark using something called Linear Temporal Logic over finite traces to evaluate whether safety properties held throughout an entire execution, not just at the end state. They define eight safety categories, things like grasp stability, cross-contamination, release stability, object containment. Real concerns, the kind that matter when you're running robots near people or near sensitive materials.

They tested six vision-language-action policies, including some well-regarded ones like pi-zero and GR00T, across 50 household manipulation tasks. The results aren't flattering. Strong models still behave unsafely in meaningful numbers of runs. And here's the part that should give pause: task-success gains did not reliably translate into safer execution. You can improve a model's ability to complete tasks and still not make it safer. Those two things are, apparently, somewhat independent.

I called my old colleague Dave, who's been doing safety validation work on collaborative robots for the better part of a decade, and his reaction was basically "yes, obviously, but it's good someone's formalizing it." That's about right. The industry has leaned heavily on task completion as the primary metric because it's easy to measure and easy to sell. Safety failures that happen mid-execution, a gripper releasing too early, a surface contaminated by a prior action, these are harder to catch and harder to quantify. SafeManip is at least giving researchers a reusable layer to diagnose them.

This is based on benchmark evaluations in simulation and controlled real-robot settings, so how all of this translates to messy real-world deployments is still an open question. But the direction is the right one.

Look, here's the thing. Neither of these papers is going to ship a product. They're not going to show up in a press release next quarter. But the problems they're addressing, robots that can't reason about their own failures, robots that complete tasks unsafely, these are exactly the problems that cause expensive incidents in real facilities. The academic work tends to lead the commercial work by a few years. Worth keeping an eye on.

James Chen · 2 days ago · 5 min