Real-time Routing under Partial Observability: Information-Efficient Policies for Connected Vehicles

ICLR 2026 Conference Submission14670 Authors

Keywords: intelligent transportation

Abstract: Real-time navigation in urban road networks requires making sequential routing decisions with incomplete and noisy information. Recent advances in IoT infrastructure and vehicle-to-everything (V2X) technologies enable connected vehicles to communicate with roadside units and traffic signals in real time. However, in practice, communication bandwidth and deployment budgets severely restrict the number of intersections that can be queried at each decision step, creating a partially observable environment for real-time navigation. Existing pipelines which separately train predictors of traffic states and then apply non-differentiable routing solvers struggle under such conditions, as they assume access to dense and complete sensing. In this paper, we present an end-to-end differentiable framework that jointly addresses vehicle-to-infrastructure(V2I) information acquisition, traffic state inference, and dynamic routing optimization. In the proposed framework, a learnable selection module proactively determines which intersections to query under communication constraints, followed by a spatio-temporal aware encoder that infers network-wide travel costs from the resulting sparse signals, and a differentiable soft shortest-path decision decoder computes re-routing strategies while allowing gradients of downstream travel cost to flow back through the entire pipeline. This tight coupling aligns model training with the true system objective of minimizing vehicle travel time. Experiments on microscopic simulation with city-scale networks demonstrate that our approach outperforms comparable baselines in travel efficiency while requiring only minimal communication. By integrating selective information acquisition and differentiable decision-making, our framework advances real-time urban navigation under partial observability and provides a scalable path toward deployment in intelligent transportation systems.

Primary Area: learning on time series and dynamical systems

Submission Number: 14670