RT-SAFE reveals safety failures hidden by navigation success
In plain words
RT-SAFE is a simulated city test for checking whether computer systems that navigate can stay safe while the world moves. It records collisions, hazard contacts, and traffic-rule violations while a system is thinking and while it carries out a move.
In tests of eight systems on difficult routes, reaching the destination stayed common when the world moved, but finishing without any safety event was rare. The test shows why arrival alone does not tell us whether a journey was safe.
The research
If an embodied agent reaches its destination, does that mean its journey was safe? In a changing environment, pedestrians, vehicles, and signals can move while the agent decides, so an action based on its last view may no longer fit the scene when it begins. This makes arrival alone an incomplete measure of safe navigation.
RT-SAFE is a simulated urban-navigation test that records safety events while agents decide and act, exposing risks hidden by task completion.
On matched hard routes, researchers compared eight vision-language models with the world paused during decisions or moving throughout decisions and actions. Task completion stayed high, but safe completion fell from 19.8% to 0.7%, and collisions rose from 3.31 to 40.68 per episode.
Most real-time collisions occurred while models were still deciding, supporting the concern that task success hides exposure to a changing scene.
The static and real-time comparison changes both simulator timing and the instructions given to models, so it cannot isolate timing as the sole cause. Hazard counts may also miss hazards crossed during movement if the agent is no longer overlapping them when the simulator checks.