Learning Direct Control Policies with Flow Matching for Autonomous Driving

Researchers introduced a flow-matching driving planner that maps bird’s-eye-view inputs directly to acceleration and curvature sequences.

Trained on simulated urban driving data from Parma, the model showed stable closed-loop performance in both familiar and out-of-distribution road scenarios.

The approach points to a faster, more robust way to generate real-time control policies for autonomous driving, with better transfer to unseen environments.