TraversRL: Traversable Pedestrian Pathway Generation With Reinforcement Learning

Traditional pedestrian mapping often relies on pixel-level segmentation, which frequently results in fragmented, non-routable graphs. This paper asks whether modeling pathway extraction as an iterative, sequential decision process—simulating a traveler navigating an environment—can produce more reliable, connected networks than static segmentation approaches.

TraversRL treats pathway generation as a reinforcement learning (RL) problem. The model uses a vision-conditioned policy to iteratively grow a graph from an aerial image, choosing from a discretized set of direction and distance actions. The training process occurs in two stages: first, supervised pretraining to imitate ground-truth graph expansions; second, RL fine-tuning using a Group-Relative Policy Optimization (GRPO) objective. The authors introduce two reward designs: a global graph-level reward based on buffered Intersection-over-Union (IoU) to ensure overall connectivity, and a local stepwise reward that provides fine-grained geometric feedback to improve edge placement.

TraversRL significantly outperforms state-of-the-art segmentation baselines across multiple datasets. The global reward model improves buffered IoU by 43.6% and connectivity metrics by 205% compared to existing methods. Furthermore, the addition of the local stepwise reward provides a consistent boost in performance, yielding cleaner graphs with fewer spurious branches. The model also demonstrates strong zero-shot generalization to unseen cities, confirming that the learned policy is robust to diverse urban layouts.

Accurate pedestrian maps are critical for urban planning and accessibility. By shifting from pixel-labeling to a topological, agent-based generation framework, TraversRL produces graphs that are inherently better suited for routing and navigation, effectively bridging the gap between raw aerial imagery and usable infrastructure data.