A human-like speed planning method for autonomous driving at roundabouts

Roundabouts require autonomous vehicles to adapt their speed to roadway geometry and nearby traffic. This study presents a longitudinal speed-planning method based on a modified Social Force Model (SFM) that reproduces speed-profile characteristics observed in human-driven roundabout trajectories. The driving-force component combines a geometry-conditioned desired-speed profile with a bounded harmonic term, and the repulsive-force component accounts for direction-dependent interaction with a nearby vehicle. The model was calibrated by nonlinear least squares using 336 retained trajectories from the rounD naturalistic driving dataset. Relative to calibrated constant-speed baselines, the in-sample Root Mean Square Error (RMSE) decreased by up to 65.5% (to 0.698 m/s) for straight maneuvers. The corresponding reduction for left-turn maneuvers was 64.0% (to 0.413 m/s). These results show improved agreement with observed longitudinal speed profiles at the studied site, while broader behavioral validity and cross-site generalization remain to be established.