LE-PAVD hybrid model slashes errors for high-speed autonomous racing

Researchers from arXiv have introduced LE-PAVD (Learning-Enhanced Physics-Aware Vehicle Dynamics), a hybrid model designed to improve high-speed autonomous navigation, particularly in racing scenarios where vehicles operate at their handling limits. Published on May 8, 2026, the paper combines physics-based priors with learned components to capture nonlinear vehicle dynamics more accurately than purely model-based or purely learned approaches. The architecture adds four key components: load-sensitive Pacejka tire forces, longitudinal load transfer, lateral tire-force effects, and rate-limited actuator inputs. Trained end-to-end on both simulation and real-world telemetry data, LE-PAVD enforces physical consistency while delivering significant improvements in state prediction accuracy.

Compared to a deep dynamics baseline, LE-PAVD reduces average displacement error (ADE) by 16.1%, final displacement error (FDE) by 20.6%, and lowers yaw-rate root mean squared error (RMSE) by a striking 91.3%. It also uses 21.6% fewer FLOPs and achieves approximately 1.50x faster inference. In closed-loop simulations, LE-PAVD achieves 17.4% faster lap times on a training track and 9.5% faster on a test track, without any track boundary violations. This compact, physics-grounded dynamics backbone offers better predictive fidelity and closed-loop performance while reducing computational cost, making it a promising approach for autonomous racing and other high-performance vehicle control tasks.