Emerging applications in autonomy are increasingly interconnected – typically consisting of many interacting agents – as in multi-robot disaster response, fleets of autonomous taxis, and automated airport ground control. Yet today's technology still struggles with the complexity of such multi-agent systems (MASs), limiting their reliability in real-world settings. Indeed, most existing MASs are still controlled, or at the very least supervised, by human operators in a ground station. Existing research has shown that guiding multiple agents to complete a mission in stressful situations can place undue cognitive burden on human operators. This increases the risk of human error that can cause loss of agents or failure of the mission, ultimately making automation essential.

The fields of formal methods, systems & control theory, and optimization have traditionally focused on system design that enjoys rigorous and verifiable safety and performance guarantees. However, these face limitations when it comes to scalability, dealing with the sophistication of AI-enabled and high-dimensional MASs, and complex system requirements, such as temporal logic or natural language requirements. On the other hand, the recent focus in machine/robot learning and AI has been on algorithmic advancements and high-performance computing. While these advances have significantly improved the performance of control systems, the integration of AI components into safety-critical settings remains largely ad hoc, with numerous instances of safety violations arising from the lack of rigorous certification.

To bridge this gap, a community has formed that works on neurosymbolic systems to combine highly performing AI models (such as neural network representations) with verifiable rule-based processing techniques (also called symbolic methods) to improve safety, explainability, and performance of autonomous systems. The goal here is to provide better computational tools for integrated safety reasoning of AI systems. To this end, in neurosymbolic systems, neural network models (or other AI models) and logical reasoning techniques are studied together as integrated models of computation. Roughly, there are three levels of operation: (1) a neural implementation of a logic, (2) a logical characterization of a neural system, and (3) a hybrid system combining a neural model and a logical characterization.

While these advances present exciting opportunities towards building efficient and performant methods to build verifiable AI systems, it also raises new questions and introduces novel challenges. This Dagstuhl Seminar on "Learning, Logic, and Control: Neurosymbolic Methods for Multi-Agent Systems" aims to unravel these challenges with a particular focus on MASs where obtaining rigorous guarantees is complicated by coupled requirements, high dimensionality, and complex agent interactions. Specifically, the seminar will address neurosymbolic verification for MASs, statistical tools for learning-enabled MAS, design principles for neurosymbolic control, control of perception-enabled systems, logic-guided and neurosymbolic learning, distributed MAS in unknown environments, and end-to-end safety in autonomy. By bringing together experts from different fields, the seminar will facilitate interdisciplinary collaborations and foster the development of new approaches that combine computational models and symbolic reasoning techniques to ensure the safe and effective use of AI technologies in MASs.

Creative Commons BY 4.0

Lars Lindemann, Laura Nenzi, Dimitra Panagou, and Jyotirmoy Deshmukh