The Robot Navigation Problem Nobody Wants to Talk About: Uncertainty

They're calling this "Multiverse Decision Making" which, call me old-fashioned, sounds like marketing got involved. But the underlying idea is sound! If your robot isn't sure whether that blob is a trash can or a person, plan routes that work reasonably well in both scenarios rather than betting everything on your first guess.

The numbers are actually impressive, though I've learned to be skeptical of benchmark results. PSG-Nav claims 66.1% success rate on MP3D, 44.8% on HM3D, and 67.9% on HSSD. For context, these are standard indoor navigation benchmarks, and those numbers represent state-of-the-art performance. The 44.8% on HM3D might look underwhelming, but that benchmark is notoriously difficult, lots of cluttered residential environments where everything looks like everything else.

The other paper, HSAN, takes a different approach to the same problem. Where PSG-Nav is all about maintaining uncertainty through the perception pipeline, HSAN focuses on hierarchical reasoning, building scene graphs that organize environments from individual objects up through regions to entire zones. Think of it as giving the robot a mental model that goes "okay, I'm in a kitchen zone, which contains a cooking region, which has a stove object." The theory is that even if you're uncertain about specific objects, you can still reason about where you probably are based on context.

They've also thrown in optimal transport theory for path planning, which is, look, I'm not going to pretend I fully understand the Kantorovich duality stuff. What I do understand is that it's trying to balance "is this the right direction semantically" against "can I actually get there physically" in a mathematically principled way. Most planners just pick the closest matching landmark and hope for the best. This tries to be smarter about it.

What strikes me about both papers is that they're finally treating uncertainty as a feature rather than a bug. For years the field has been obsessed with making perception more confident, more accurate, more certain. These approaches flip that around and say, okay, what if we just accept that our robots are going to be confused sometimes and build systems that handle confusion gracefully?

This matters more than you might think. Open-vocabulary navigation, where robots have to find objects they weren't specifically trained on, is becoming table stakes for real-world deployment. Your warehouse robot needs to find "the blue bin by the loading dock" even if it's never seen that specific bin before. Your home assistant needs to fetch "the thing I left on the couch" even when your couch is covered in laundry (not that I would know anything about that).

The PSG-Nav paper includes something called an "Evidential Experience Calibrator" which is basically the robot learning from its own mistakes over time. If it keeps misidentifying certain objects, it adjusts its confidence accordingly. This is the kind of online learning that makes me nervous, there are failure modes here that the paper doesn't fully explore, but it's also probably necessary for real deployment.

I should note what we don't know yet. Both papers are evaluated on simulation benchmarks, not real robots in real environments. The gap between simulation performance and real-world performance in robotics is, historically, brutal. Lighting changes, sensor noise, objects that don't exist in your training distribution, all of this eats into those nice benchmark numbers. The PSG-Nav team does provide code, which is good, but I haven't seen real-world validation data.

There's also the compute question. Maintaining full probability distributions and sampling multiple possible worlds isn't free. The papers are light on inference time details, which usually means it's not great. For a research paper that's fine, for a robot that needs to make decisions in real-time while navigating around humans, it might be a dealbreaker.

Still, I'm cautiously optimistic here. The self-driving car industry spent a decade trying to eliminate uncertainty through better sensors and more data, and it turns out the real world is just irreducibly uncertain. Robotics seems to be learning that lesson faster. These papers represent a philosophical shift toward embracing uncertainty rather than fighting it, and that feels like progress to me.

Whether it actually translates to robots that work better in your home or warehouse remains unclear. But the direction is right. If you want to argue about it, my email's on the about page.

Robert "Bob" Macintosh · 3 hours ago · 4 min