Generating accurate long-range predictions about the behavior of other agents is hard enough; getting them to not collide and obey the rules of the road nearly all the time is much harder.

Inverted AI has developed a technique for "fine-tuning" ITRA to novel intersection geometries such that infractions are significantly minimized.

For example when we train ITRA-0 on the non-commercial publicly available Interpret challenge Interaction data our agents stay on road and don't collide. But when we ask our ITRA-0 agents to drive in new locations, for instance in CARLA simulator locations as below, we see increased infractions largely due to domain (distribution) shift in the road layout.

ITRA-1, aka TITRATED, addresses this issue resulting in a 70% reduction in infraction rates. Examples of ITRA-1 driving in CARLA are shown in the video above. Note that ITRA-1 drives all cars.

Read the TITRATED workshop paper for more details or watch the talk below for highlights.