Current autonomous vehicles are typically deployed within limited geographic regions, while scalable operation across diverse locations is increasingly demanded. As deployment regions expand, planners must cope with heterogeneous traffic dynamics under fixed on-board computational and memory budgets, where adapting a single monolithic model through data aggregation or parameter expansion becomes inefficient and costly. This paper proposes the Dynamically Local-Enhancement (DLE) planner, which improves region-level adaptability without increasing policy capacity or performing full online policy optimization during deployment. Global driving competence is decoupled from region-specific adaptation through explicit region-conditioned representations. Long-term regional characteristics are distilled into map-level historical memory via a latent-variable encoder, while real-time interactions are modeled by a dual-layer traffic graph neural network that jointly captures vehicle interactions and road topology. The resulting region-conditioned representation is used to condition a shared reinforcement learning planner at inference time, where dynamic behavior arises from location-indexed retrieval and conditional forward inference rather than parameter growth. We evaluate DLE in multi-region closed-loop CARLA benchmarks. Under a fixed parameter budget, DLE consistently improves cross-regional adaptability and outperforms baselines in safety and comfort metrics. These results indicate that memory-based region-aware enhancement offers a practical paradigm for scaling autonomous driving planners under deployment constraints.
Deng et al. (2026) studied this question.