Robots That Know When to Leave Things Alone: Two New Papers Worth Your Time

They also introduce a new benchmark dataset called APOR, which is synthetic but designed to capture messier, more realistic conditions: cluttered rooms, noisy observations, multiple furniture items, and explicit abstention cases. It's too early to say how well this generalises beyond the controlled settings they tested, but the initial results are promising.

The second paper is about something slightly different but related. It's called HAVE, short for History-Aware VErifier, and the full paper is at arXiv cs.RO. The problem here is visually ambiguous objects, the kind where you can't tell from a single observation how something will behave when you interact with it. Think doors that might be push or pull, objects with uneven weight distribution, that sort of thing.

The approach decouples action generation from verification. A diffusion-based generator proposes multiple candidate actions, and then the verifier looks at the robot's history of past interactions with that object (or similar ones) to pick the best candidate. The theoretical argument is clean: having a separate verification step improves expected action quality, and the empirical results across both simulated and real-world tests back that up. The project site is at https://liy1shu.github.io/HAVE_CoRL25/ if you want to dig into the videos.

What I like about both of these, and this is the thing that connects them, is that they're both sort of accepting that robots will encounter situations they can't handle cleanly. That's a more mature framing than a lot of work I've seen. The goal isn't omniscience. It's knowing your own limits.

Now, the caveats. Both papers are academic work with synthetic or lab-scale evaluations. The APOLLO dataset is LLM-generated, which introduces its own biases. HAVE tests on articulated objects and multi-modal doors, which are interesting but a long way from a busy warehouse floor. I only found these two sources on this particular direction of research, so I can't tell you whether this is a growing trend or two teams independently stumbling onto the same good idea.

What I can say is that the underlying engineering instinct is right. Uncertainty quantification, abstention, history-aware decision-making. These are the kinds of things that separate a robot that works in a demo from one that works in the field. Anyone who's spent time commissioning real systems knows the difference.

I'll be keeping an eye on both of these.

The HPE AI Factory expansion announced at HPE Discover isn't just a server upgrade. It's a signal that agentic AI is moving out of the lab and into production environments.

Robert "Bob" Macintosh · 21 hours ago · 4 min