Scenario orchestrator keeps self-driving tests on script when the car under test reacts

Researchers led by Frieda Rong, with Raquel Urtasun and Igor Gilitschenski among the authors, have described in an arXiv paper a scenario orchestrator for testing self-driving policies in simulation. Scenarios scripted in advance, with fixed routes and speeds for other road users, can lose the interaction they were meant to test once the car under test reacts.

The orchestrator watches the whole traffic situation and adjusts the other actors by changing their timing, routes or roles, solving a constrained optimisation at each step to keep the intended interaction while limiting interventions. It also separates a scenario that failed to happen from a driving policy that failed. In the CARLA simulator, it produced the targeted interactions more reliably than scripted and learned baselines, especially when starting conditions and the tested car's behaviour varied.