Flow matching generates steering commands directly for autonomous vehicles

Autonomous driving systems typically decompose the problem into perception, planning, and control. This work trains a single neural network to output acceleration and curvature trajectories directly from bird's-eye-view scene representations. The flow-matching formulation generates control sequences via a few ODE integration steps, enabling low-latency closed-loop re-planning. Trained exclusively on urban simulator data from Parma, Italy, the model generalizes to both out-of-distribution highways and unseen urban scenarios. The authors attribute this robustness to the geometry-centric BEV representation and the smooth vector field learned by flow matching, which degrades gracefully under distribution shift. Closed-loop experiments confirm stable control across diverse environments.