LLM-Guided Safety Agent for Edge Robotics with an ISO-Compliant Perception-Compute-Control Architecture

A research team led by Xu Huang has published a paper detailing a novel LLM-Guided Safety Agent designed to solve a core conflict in robotics: the probabilistic nature of AI perception versus the deterministic behavior required by industrial safety standards like ISO 13849. Their architecture uses large language models (LLMs) as a bridge, translating written safety regulations and natural-language commands into formal, executable logic predicates. This allows safety rules to be updated and deployed without deep code-level changes, making the system more adaptable.

The system's physical implementation is built for the constraints of edge computing, where robots operate with limited power and compute. It employs a symmetric dual-modular redundancy design, meaning two independent compute modules (prototyped on a dual-RK3588 platform) run in parallel for fault tolerance. This ensures low-latency perception, computation, and control in a closed loop, meeting the high-reliability requirements for Performance Level d (PL d) and Category 3 safety. Crucially, it demonstrates this high safety level using affordable, off-the-shelf hardware, paving a practical deployment path for safety-critical embodied AI in manufacturing, logistics, and healthcare where cost has been a barrier.