Dynamically Local-Enhancement Planner for Large-Scale Autonomous Driving

TL;DR

This work tackles the challenge of scaling autonomous driving policies across diverse regions by introducing the Dynamically Local-Enhancement (DLE) planner, which augments a base policy with local regional data without permanently enlarging the model. It formalizes a Position-Varying MDP (POVMDP) and uses a Graph Neural Network to extract region-specific features from local observations, storing them in a Regional Data Container and feeding them into a dynamic policy enhancement stage guided by mutual-information objectives. The approach demonstrates that DLE achieves higher safety (lower collision rates) and better average rewards than single-model baselines and approaches the performance of a large global model, while maintaining a lighter computational footprint suitable for large-scale deployment. Overall, DLE offers a scalable pathway for cross-regional autonomous driving, enabling region-aware decision-making without substantial increases in on-device model size or training burden.

Abstract

Current autonomous vehicles operate primarily within limited regions, but there is increasing demand for broader applications. However, as models scale, their limited capacity becomes a significant challenge for adapting to novel scenarios. It is increasingly difficult to improve models for new situations using a single monolithic model. To address this issue, we introduce the concept of dynamically enhancing a basic driving planner with local driving data, without permanently modifying the planner itself. This approach, termed the Dynamically Local-Enhancement (DLE) Planner, aims to improve the scalability of autonomous driving systems without significantly expanding the planner's size. Our approach introduces a position-varying Markov Decision Process formulation coupled with a graph neural network that extracts region-specific driving features from local observation data. The learned features describe the local behavior of the surrounding objects, which is then leveraged to enhance a basic reinforcement learning-based policy. We evaluated our approach in multiple scenarios and compared it with a one-for-all driving model. The results show that our method outperforms the baseline policy in both safety (collision rate) and average reward, while maintaining a lighter scale. This approach has the potential to benefit large-scale autonomous vehicles without the need for largely expanding on-device driving models.

Figures (7)

• Figure 1: Main idea: Dynamically enhancing a basic driving policy with local driving data when the autonomous vehicle driving to different regions, improving the scalability of the system without significantly expanding the planner’s size.
• Figure 2: Our framework establishes a hybrid policy architecture that enhance a basic policy with dynamic local adaptation: The basic policy $\pi_b$ works globally, while parallelly, a local-enhancement module extracts region-specific features $g$. Finally, the basic policy will be updated dynamically via information-theoretic optimization. This dual-stream design enables local adaption of driving policy through historical driving data collected locally.
• Figure 3: Connection relationships between local feature nodes. Road nodes connect to previous road nodes, and vehicle nodes connect through nodes of interest.
• Figure 4: Local information extraction process.
• Figure 5: The test scenarios sourced from different regions are designed to have distinct road structures and surrounding vehicle behaviors. In the upper scenario, the merging - in vehicle is likely to maintain its lane. Conversely, in the lower scenario, it is more prone to changing lanes.