The adoption of neural networks for motion control modules emerges as a critical direction in the advancement of end-to-end autonomous driving. However, few studies have comprehensively addressed the challenges of robustness and generalization in motion control policies, including long-tailed distribution, distribution shift, and sim-to-real gap. In practical applications, motion control performance is compromised by diverse uncertainties, posing substantial challenges to real-world deployment. This work develops a training system to enhance the robustness and generalization of motion control policies when passing through multiple intersections. We first construct a task library comprising 6 driving scenarios, which are allocated to different sampling processes to rebalance the proportion of monotonous and edge scenarios. Next, we formulate a zero-sum game for uncertainties and driving actions with smoothing constraints within the range of observation noise. The driving policy is optimized by the proposed robust policy iteration method for the worst-case performance, which is approximated via Taylor expansion to avoid the computational burden caused by adversarial training on behavior disturbance, where the approximate results decouple model mismatches to ensure robust performance and action smoothness is boosted through penalty function method. Ultimately, the motion control performance and the robustness of driving policy are thoroughly validated by configuring the behavior patterns of traffic participants, ego dynamic parameters, and observation noise intensities in the simulation environment. Physical vehicle experiments on public urban roads further depict the robustness and generalization of the driving policy learned from simulations.

Controllability is a fundamental property of control systems, serving as the prerequisite for controller design. While controllability test is well established in modelic (i.e., model-driven) control systems, extending it to datatic (i.e., data-driven) control systems is still a challenging task due to the absence of system models. In this study, we propose a general controllability test method for nonlinear systems with datatic description, where the system behaviors are merely described by data. In this situation, the state transition information of a dynamic system is available only at a limited number of data points, leaving the behaviors beyond these points unknown. Different from traditional exact controllability, we introduce a new concept called ϵ-controllability, which extends the definition from point-to-point form to point-to-region form. Accordingly, our focus shifts to checking whether the system state can be steered to a closed state ball centered on the target state, rather than exactly at that target state. Given a known state transition sample, the Lipschitz continuity assumption restricts the one-step transition of all the points in a state ball to a small neighborhood of the subsequent state. This property is referred to as one-step controllability backpropagation, i.e., if the states within this neighborhood are ϵ-controllable, those within the state ball are also ϵ-controllable. On its basis, we propose a tree search algorithm called maximum expansion of controllable subset (MECS) to identify controllable states in the dataset. Starting with a specific target state, our algorithm can iteratively propagate controllability from a known state ball to a new one. This iterative process gradually enlarges the ϵ-controllable subset by incorporating new controllable balls until all ϵ-controllable states are searched. Besides, a simplified version of MECS is proposed by solving a special shortest path problem, called Floyd expansion with radius fixed (FERF). FERF maintains a fixed radius of all controllable balls based on a mutual controllability assumption of neighboring states. The effectiveness of our method is validated in three datatic control systems whose dynamic behaviors are described by sampled data.