Autonomous-driving tests stop following a fixed script: a scenario orchestrator rewrites other vehicles in real time and fully realizes the intended interaction in all six scenario families
Synopsis
Autonomous-driving policies react in closed loop, so a scenario scripted with fixed routes often drifts away from the interaction it was meant to stage, a problem called closed-loop drift. A scenario orchestrator continuously monitors the joint traffic state and adapts non-ego actors through re-timing, re-routing and role reassignment, solving a constrained optimization at each step so the test intent survives while interventions stay limited and traffic stays feasible. In the CARLA simulator across six scenario families and six ego policies, the method fully realizes the intended interaction in every scenario family, while baselines such as ProSim, LASER and OSC2Runner mostly satisfy only the two easiest scenes.
Interpretation
Test scenarios are no longer fully fixed before execution; an orchestrator keeps adjusting non-ego vehicles during simulation so the intended interaction still occurs once the ego policy responds in real time. Scripted execution fixes the scenario definition at run time, leaving the intent open-loop, while data-driven learned actors are reactive but favor nominal, in-distribution behavior and rarely realize a specific safety-critical interaction. In CARLA, six scenario families and six ego policies form 144 evaluation cells; the orchestration method satisfies all three scenes in all six scenario families, whereas baselines such as ProSim, LASER and OSC2Runner mostly satisfy only the two easiest scenes.
The orchestrator treats a scenario as a set of high-level directives, checks each step whether they still hold, and intervenes to adjust vehicle behavior when they do not. It has three intervention types: re-timing adjusts a vehicle's speed to match the ego's predicted occupancy window at the conflict point, re-routing rebuilds a vehicle's remaining maneuver sequence when it is not mid-turn, and role reassignment switches the hero vehicle when the new candidate has a cheaper intervention cost. Four directives drive these decisions: whether any actor collides with the ego within the prediction horizon, whether the hero can still reach the conflict point, whether a third actor blocks or collides first, and whether background actors collide with each other; all are re-evaluated at every timestep.
Perspective
The demonstrated scope is six scenario families and six ego policies in the CARLA simulator, with scenario intent set by manually configured lanes, ego route and available actions; the authors list combining data-driven actor models with the existing objectives and constraints as future work.
The authors state that the current design relies on manually specified traffic behaviors, such as ranking actors by their likelihood of carrying out the intended action, which will not scale as scenarios become more complex and open-ended; the game tree is currently a chain of sequential nodes, with branching search left to future work.