Predictive control for dynamic drifting path tracking in autonomous vehicles with deep Koopman hybrid modeling and adaptive mode switching

Professional race drivers exploit vehicle handling limits via controlled drifting for rapid cornering; endowing autonomous vehicles with this capability could markedly enhance extreme-condition performance and driving safety. However, existing strategies struggle to integrate drifting and normal steering for cooperative path tracking: they fail to determine optimal switching timings for both normal and extreme cornering, while the strong inherent nonlinearity of drifting poses challenges to real-time optimal control. To address these gaps, this paper proposes a hierarchical coordinated drift-normal path-tracking strategy with deep reinforcement learning (DRL)-based mode switching, achieving superior tracking accuracy, stability, and computational efficiency. The upper layer adopts a heuristic-rules-embedded DRL policy to optimize drift-normal transitions by refining mode switching timings. These rules ensure DRL intervention to proper corner entry and exit phases, improving switching efficacy and mitigates instability risks from potential overgeneralization. The lower layer implements mode-specific motion control: normal driving uses a proportional-integral-derivative controller and linear quadratic regulator, while drifting is controlled by a computationally efficient Koopman-based predictive controller (KMPC) built on a linearized hybrid model integrating an inverse tire model and deep Koopman operator. Validation results confirm that the proposed strategy improves path tracking precision and driving stability via well-timed mode transitions and superior motion control. The proposed KMPC delivers comparable performance to nonlinear model predictive control while cutting maximum single-step computation time by 85.72%. The proposed strategy also exhibits favorable generalization to untrained complex paths and strong robustness against mismatched tire-road friction and vehicle mass parameters.