CAPE framework benchmarks racing controllers with 50% faster lap times

Researchers from the CAPE project (Malik Ali, Musabbir Ahmed Arrafi, Nicholas M. Stiffler, Krishna Bhavithavya Kidambi) introduce the Control Algorithm Performance Evaluation (CAPE) framework, a systematic methodology for benchmarking racing controllers under learned vehicle dynamics. Their key innovation is the Enhanced Physics Model (EPM), which combines physical modeling with learned corrections. In closed-loop experiments comparing EPM against two state-of-the-art models—Deep Pacejka Model (DPM) and Deep-learning Dynamics Model (DDM)—the team tested five control architectures. Adaptive NMPC with EPM achieved a best average lap time of 5.82 seconds on a simulated track, dramatically outperforming DPM (12.99s) and DDM (8.80s). EPM also produced substantially lower longitudinal and lateral tracking errors under identical controller settings.

The researchers further evaluated all models under a disturbance-aware simulation incorporating measurement noise, process disturbances, actuator delay, and parametric uncertainty. With a moderate global disturbance scaling factor (η = 1), results averaged across five controllers show EPM reduces longitudinal tracking error by 29.0% (vs DPM) and 17.2% (vs DDM), and lateral tracking error by 24.6% and 12.3%, respectively. Simultaneously, average velocity magnitude increased by 39.9% over DPM and 3.1% over DDM. CAPE thus establishes a systematic benchmark for evaluating learned vehicle dynamics in closed-loop control, demonstrating that EPM significantly improves controller robustness and performance under realistic uncertainties. This work has direct implications for autonomous racing, advanced driver assistance, and any application requiring high-performance vehicle control with learned models.