SafeDriveVLA overrides unsafe navigation commands in simulated driving

A coherent navigation instruction can conflict with the environment in which a vehicle must act. SafeDriveVLA gives a vision-language-action controller an explicit mode for strict compliance, cautious compliance or fallback, and uses a navigation-conditioned world model to estimate consequences before choosing maneuvers. In simulations, the model may brake or steer around a newly detected obstruction rather than continuing to obey a speed or lane request literally.

In the researchers' conflict benchmark B2D-C, SafeDriveVLA reported a Driving Score of 85.9, success rate 67.3%, 38 collisions, nine traffic violations and four out-of-route events. The selected SimLingo-Safe baseline recorded score 72.8, success 38.0%, 60 collisions, 61 violations and 11 out-of-route events. These are simulator counts, not traffic accidents on public roads. Training and evaluation choices shape both obedience and safety scores.

The work identifies a transferable authority distinction: a user's requested objective is not automatically a permitted physical effect. Its limitations matter as much as the apparent improvement. A learned world model can fail to predict hazards, and unjustified overrides can also create risk. No real-world safety certification or independently replicated road deployment follows from the reported scores.