Background and motivation
Aging road infrastructure increasingly compromises safety and economic performance. Robots can already inspect roads and repair selected defects, but current systems remain largely semi-autonomous and dependent on human supervision. Established maintenance management procedures support decisions at network or section level, yet translating selected measures into robotic operations at work-zone level remains difficult. Robust scene understanding is therefore a key prerequisite for higher autonomy in robot-based road infrastructure maintenance.
Research objectives
The project aims to develop a methodology for scene understanding tailored to robot-based road infrastructure maintenance, focusing on work-zone-level execution rather than maintenance management at network or section level. The project does not develop a full road-repair machine or asphalt-processing hardware. Instead, the project provides the methodological basis that existing and future maintenance robots require to move from semi-autonomous task execution toward higher levels of autonomy. The underlying hypothesis is that scene understanding improves robot-based road infrastructure maintenance. To achieve the objective, robot capabilities and maintenance procedures will be formalized in a mathematically sound knowledge representation, and digital road models will be exploited systematically. Scene graphs will capture maintenance-relevant semantics and hierarchies, as shown in Figure 1. A framework for scene understanding will generate and fuse ontology-grounded multi-view scene graphs from visual data and digital road models and enrich incomplete scene representations under partial observability. A framework for maintenance planning will couple multi-view scene graphs with planning languages, large language models, and temporal logic to synthesize coordinated maintenance plans and monitor correct execution. The project therefore addresses methodological challenges in formal knowledge representation, multi-view scene graph fusion, knowledge enrichment, planning integration, and runtime monitoring. Validation will be carried out in a staged manner through simulations, laboratory experiments, and field tests of selected core functions in cooperation with public road operators, who will provide infrastructure data, planning documents, and access to real-world operational environments.
Expected outcome
The expected outcome is a validated methodology for more transparent, reliable, and safety-aware robot-based road infrastructure maintenance. Beyond the application domain, the project is expected to contribute generalizable methodological advances in knowledge-based scene representation, formal planning, semantic reasoning, and execution monitoring for autonomous robotic systems.
