Answers are grounded in this paper and cite their sources.
This paper addresses lateral path tracking for an autonomous vehicle whose state-space matrices are unknown. The infinite-horizon linear-quadratic (LQ) optimal steering gain is learned from a single batch of measured state and steering data by off-policy policy iteration, which needs neither the system matrices, nor state derivatives, nor a function-approximation network; each iteration is a linear least-squares problem built from three reusable data integrals, and the iteration coincides with Kleinman’s model-based iteration under a rank condition that we prove in the symmetric quadratic basis. Because the deployed controller is a fixed linear gain, we do not claim that a barrier transformation enforces the tracking bounds. Instead we derive a prescribed-performance certificate : an explicit sufficient per-channel exponential upper bound obtained from the modal decomposition of the closed loop, together with sufficient conditions relating the funnel parameters to the closed-loop decay margin and to the (unmatched) road-curvature disturbance, and a companion bound on the steering angle. Simulations on a 30 km/h S-turn and an 80 km/h lane change show that the learned gain is within 2–6% of the Riccati gain and within \(3× 10⁻⁶\) (relative) of its nominal LQ cost, and that all errors remain inside the certified funnels. We also report the limits of the approach: the certificate is nominal and covers 200/200 and 198/200 trials of a Monte-Carlo study with \(± 20%\) speed and \(± 30%\) cornering-stiffness variation; the learner is markedly more sensitive to measurement noise than identification followed by a Riccati solve; and large initial errors saturate the steering actuator, which requires retuning at the price of a wider certified funnel.
No takes yet. Share an insight, caveat, or question.
Lu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: