Researchers have built a self-driving AI model that predicts not just what the road looks like, but where it is actually safe to go.
The model, called AffordDrive3D, pairs a vision-language-model backbone with a world-action model that forecasts future RGB scenes and 3D geometry together. The twist is an added affordance layer that flags drivable areas and collision-critical zones tied directly to the vehicle's next move, rather than just mapping the whole scene's shape. On the NAVSIM benchmark, the system scored 91.3 PDMS and 89.9 EPDMS, which the researchers describe as state-of-the-art.
Most prior world-action models treat geometry as one flat layer - useful for knowing the shape of the world, but silent on which parts of that shape matter for a given maneuver. By scoring affordance and geometry together, AffordDrive3D gives the planning layer a shortlist of relevant risks instead of a full scene to sort through, which is the kind of distinction that tends to separate benchmark wins from road-ready systems.
NAVSIM scores are a useful shorthand, but they are still simulation numbers - the real test is whether these affordance maps hold up on unpredictable streets, not clean benchmark replays.