Sbir phase i: AI-enabled affordable and autonomous road condition assessment system

The National Science Foundation awarded Pavex LLC $305,000 on July 1, 2026, under the NSF Technology, Innovation, and Partnerships program (CFDA 47.084) for a Phase I Small Business Innovation Research project to develop an AI-enabled system for autonomous road condition assessment.

The project delivers a low-cost, vehicle-mounted sensing approach that collects street-level data and produces lane-level road condition ratings to support infrastructure maintenance planning. The work addresses two operational bottlenecks: manual route planning and manual validation of distress detections. The recipient will develop a scalable route optimization architecture that combines offline graph partitioning with online deep reinforcement learning to generate traffic-aware, constraint-aware inspection routes from municipal geographic information system road files. In parallel, the project develops a Bayesian multi-frame data fusion framework that integrates redundant observations across consecutive image frames to suppress false positives, improve distress detection reliability, and reduce reliance on human review. Phase I success metrics include measurable reductions in route planning overhead and human validation effort alongside improved detection performance under real-world variability.

Performance occurs in West Lafayette, Indiana, with a period of performance from July 1, 2026, through December 31, 2027.

Abstract

The broader/commercial impact of this small business innovation research phase i project is to develop a low-cost, vehicle-mounted sensing approach that can be deployed widely to collect street-level data and produce practical road condition ratings that help agencies plan repairs sooner and more efficiently. By enabling more timely maintenance decisions, the project can reduce vehicle damage costs while improving safety. This project will investigate an automation framework that enables complete, lane-level road coverage using low-cost mobile sensors while reducing two major operational bottlenecks: manual route planning and manual validation of uncertain distress detections. The project will investigate a scalable route optimization architecture that combines offline graph partitioning with online deep reinforcement learning to generate traffic-aware, constraint-aware inspection routes that reflect real-world road networks. This project will develop methods to convert municipal geographic information system road files into directed, traffic-weighted graphs and partition them into feasible daily subgraphs. A deep reinforcement learning agent, augmented with graph neural network embeddings, will then learn efficient within-partition traversal strategies and adapt to changing conditions during deployment. In parallel, the project will improve distress detection reliability by developing a bayesian multi-frame data fusion framework that integrates redundant observations across consecutive frames. Rather than treating each image independently, the fusion approach will combine evidence over time to suppress false positives and strengthen consistent detections, improving accuracy and reducing reliance on human review. Phase i success targets include measurable reductions in route planning overhead and human validation effort, while improving detection performance under real-world variability. The resulting methods will advance artificial intelligence for infrastructure monitoring. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the foundation's intellectual merit and broader impacts review criteria.- subawards are not planned for this award.