Two Papers That Actually Get Why Robots Keep Bumbling Into People

Look, here's the thing: I've watched more robots navigate into awkward situations than I care to count. Back when I was at Kuka, we had a mobile platform demo at Automatica (2014, I think) where our AGV kept trying to split up couples walking through the booth. Technically collision-free. Socially, a disaster. The sales guys were not pleased.

So when two papers dropped recently addressing exactly this kind of problem, I paid attention.

The Group Problem Nobody Wanted to Solve

The first paper, TAGA (Tangent Action for Group Avoidance) from a team including researchers at several universities, tackles something I've been complaining about for years: robots treat humans as independent obstacles. They don't understand that three people standing in a triangle are having a conversation, not just occupying three separate points in space.

arXiv has the full paper, and what I like about it is the pragmatism. They're not trying to rebuild navigation from scratch. TAGA sits on top of your existing navigation stack and handles group avoidance through tangent-path maneuvers. You keep your collision avoidance, you add social awareness.

They also introduce something called Group Crossing Rate, which measures how often a robot barges through the middle of a social group. It's a continuous metric, not just "did you hit anyone?" This matters. I called my old colleague at Siemens last week, and he made the point that warehouse robots are increasingly sharing space with human workers, not just operating in segregated zones. The days of just avoiding collisions are ending.

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.