Researchers led by Seokjun Lee have built a planner for drifting, driving a car on purpose at a large sideslip angle, that takes the desired angle as a direct input, according to a paper on arXiv. Drifting matters in motorsport and for recovering from sudden oversteer, and the authors say theirs is the first generative drift planner conditioned on a requested peak angle.
A flow-matching model trained on demonstrated drift laps and refined with reinforcement learning generates the paths. A control barrier function then corrects them to respect the road, tyre friction and the limits of the steering, throttle and brakes.
In the Assetto Corsa racing simulator, a predictive controller followed the plans for six requested angles from 20 to 45 degrees, over 30 laps, about 75 km, for each. The measured peak angle matched the request with an average error of 1.2 degrees, and the correction step sharply reduced runs off the road and spins.



