Interaction-Conditioned Navigation for End-to-End Autonomous Driving

Reactive driving requires temporary navigation changes in response to other road users, but trajectory imitation does not explicitly supervise the navigation condition connecting an interaction to its execution. CondAD localizes candidate actors, relates them to the ego route, and verifies their identity on detection-guided image crops. The actor–route interaction becomes a navigation condition, a maneuver command together with a spatial target, and a persistent maneuver state keeps that condition across planning steps. On Bench2Drive, stacked-yield route success rises from 40.00% to 90.00% (9/10), a gain of 50 percentage points. That figure is separate from the Give Way ablation, where Full CondAD reaches 70.00% on the 10-route Give Way set. Progressive ablations on that set support converting trigger information into a command–target condition and retaining the maneuver state during execution.