Eight rival self-driving planning algorithms just took the same test, and none of them aced it.
A new comparative study pulls representative motion planners from three separate benchmark ecosystems, CARLA, nuPlan, and the Waymo Open Dataset, then reruns all of them on one shared platform: CARLA Leaderboard v2.1. The eight methods under evaluation are TF++, InterFuser, TCP, PDM-Lite, MTR+MPC, CaRL, PlanT 2.0, and Diffusion Planner. Rather than proposing a new algorithm, the researchers act as referees, scoring each system's strengths and weaknesses under identical driving scenarios and conditions. The goal is a genuinely apples-to-apples read on where planning research actually stands.
That matters because autonomous driving research is scattered across benchmarks that do not talk to each other. A planner that tops the nuPlan leaderboard was not necessarily built or tuned for CARLA's scenarios, and vice versa, so cross-ecosystem comparisons like this one are rare and useful for spotting which techniques generalize versus which just fit one dataset's quirks.
Leaderboards make for good marketing copy, but a single unified test run is still a long way from a car handling a real intersection.