Learning while teaching for autonomous driving

Abstract Autonomous driving decision-making faces significant challenges due to complex environments and dynamic uncertainties. While knowledge distillation offers potential, conventional fixed teacher–student roles and unidirectional knowledge transfer limit flexibility and efficiency. To this end, we introduce the Learning while Teaching (LwT) framework. LwT establishes a dual-model reinforcement learning structure with dynamic role switching, enabling bidirectional knowledge transmission. LwT is a dual-flexible framework, allowing the primary model to learn through three synergistic pathways: direct environment interaction, auxiliary feedback, and auxiliary guidance. Addressing key challenges of primary model expressiveness and malignant positive feedback, LwT incorporates: (i) a Tri-Group synergistic reinforcement learning for enhanced expressiveness via joint multi-group optimization, and (ii) a collaborative feedback arbiter utilizing a self-distilled Kolmogorov–Arnold network to ensure beneficial teaching-learning interactions. Extensive experiments demonstrate that LwT significantly outperforms baselines, achieving 15.68% higher performance in simulation and a 13.3% increase in real-world success rate, while effectively improving training efficiency and decision-making quality in autonomous driving systems.