Protocols that respect semantic safety constraints in learned multi-agent communication and control

Develop communication and control protocols for multi-agent systems that guarantee compliance with semantic safety constraints—implicit, context-dependent rules that extend beyond geometric state constraints—and provide methods to enforce these constraints during learning and execution of emergent communication policies.

Background

Safety filters in control often rely on geometric constraints and do not capture implicit, semantic safety rules encountered in human-centric or complex environments. As emergent communication policies become prevalent, ensuring that message-driven decisions respect such constraints is essential.

The paper highlights that designing protocols to accommodate semantic safety constraints remains unresolved, calling for methods that integrate high-level semantic reasoning with reliable communication and control.