Two Papers That Actually Get Why Robots Keep Bumbling Into People

The results show TAGA cuts group crossing roughly in half for classical reactive planners, with minimal impact on success rates. For learned policies, the gains are smaller, which makes sense. Those systems have already absorbed some social awareness from training data. The paper's honest about this, which I appreciate.

The Speeding Problem

The second paper, Plan-R1 from a team working on autonomous driving, addresses a different but related issue: learning-based planners inherit bad habits from human demonstrations. Including, apparently, speeding.

I'll be honest, this made me laugh. We spent years trying to get robots to behave more like humans, and now the problem is they behave too much like humans. Including the parts where humans cut corners and drive like they're late for something.

arXiv describes their two-stage approach. First, train on human data to get natural-looking behavior. Second, fine-tune with rule-based rewards to stamp out the bad habits while keeping the good ones. They call the second stage Group Relative Policy Optimization, or GRPO.

The clever bit is their fix for a problem with GRPO itself. Normalizing rewards across groups was washing out safety-critical signals, making rare dangerous situations look statistically similar to common safe ones. Their Variance-Decoupled GRPO preserves the absolute magnitude of rewards, so safety violations stay prominent during training.

I don't have access to their benchmark results beyond what's in the abstract, so I can't verify the "state-of-the-art" claims independently. But the approach seems sound.

What This Actually Means

Neither of these papers is going to revolutionize anything next quarter. That's not how this works. But they represent something I've been waiting to see: researchers taking social and safety constraints seriously as engineering problems, not afterthoughts.

When I started in this industry, the attitude was "don't hit anything and you're fine." Then it became "don't hit anything and don't scare people." Now it's becoming "don't hit anything, don't scare people, don't be rude, don't speed, and don't learn bad habits from the humans you're supposed to be helping."

That's progress. Messy, incremental progress, but progress.

The question I can't answer yet is how well these approaches transfer to real-world deployment. TAGA's been tested in simulation with synthetic pedestrian models. Plan-R1 uses the nuPlan benchmark, which is closer to reality but still, you know, a benchmark. We won't know for certain until someone puts these ideas into actual hardware sharing actual space with actual humans who don't read the documentation.

I've seen enough promising research die in that gap to be cautious. But I've also seen enough to know that sometimes it doesn't.

Sarah Williams · 19 Jun · 3 min