The other-driver problem

For a decade, self-driving has been a perception race: better cameras, denser point clouds, larger models. The fumbles the public actually sees — the unprotected left, the four-way stop, the two cars lunging for the same gap — are not perception failures. They are negotiations the car never entered. Call it the other-driver problem: the slice of driving where the world is shaped by what other humans are about to do.

The fix is not more resolution. It is a different unit of analysis. The DecisionPerceiver preprint, an attention-based architecture tested across three scenarios with rising interaction demand, reports consistent gains when the model treats other agents as conversation partners with their own goals rather than moving boxes to avoid. The mechanism generalizes: build a shared attention space over every agent's likely next move, and let the policy choose inside that space instead of around it. Selection story first, causal second — architectures that learn to model intent get better numbers, but models that only learn intent without a shared representation still miss the merge.

The preprint is one entry. The pattern is wider: the next self-driving gains will not come from sharper eyes, but from a car that can finally answer the question every intersection asks — what is that other driver about to do?