A cheap software patch made self-driving AI models track their own predicted paths far more reliably, without any retraining.
Researchers built a lightweight postprocessing step called Endpoint-Constrained Optimization, or ECO, that sits between an end-to-end driving model and the controller that steers the car. Most driving AI is trained by mimicking recorded routes, then it has to run in real time on the road, and that mismatch often produces trajectories that are physically awkward to follow even when the start and end points look fine. ECO keeps a model's predicted endpoint untouched and instead smooths out the intermediate waypoints so they are easier for a controller to actually execute. It needs no map data, no privileged simulator information, and no extra training, so it can be dropped onto a wide range of existing driving policies.
Applied to six different driving policies across two closed-loop simulators, ECO improved every single one, with the biggest gains going to models that most often broke physical motion limits. On the HUGSIM benchmark it took VaVAM from a score of 18.1 to 31.0, a 71 percent jump that was enough to win the HUGSIM Closed-Loop Driving Challenge, and on AlpaSim it lifted VaVAM's scene score by 123 percent. That is a striking amount of performance recovered from a step that changes nothing about how the underlying model was trained.
It is a reminder that a lot of the gap between a self-driving demo and a self-driving car lives in unglamorous plumbing, not in bigger models.