NEXUS: Continual Learning of Symbolic Constraints for Safe and Robust Embodied Planning
Researchers have introduced NEXUS, a novel modular framework designed to address the safety gap between the probabilistic nature of Large Language Models (LLMs) and the strict determinism required in physical embodied intelligence. Unlike previous approaches that treat symbolic artifacts as static interfaces, NEXUS utilizes them for symbolic grounding and continuous knowledge evolution. The framework explicitly decouples physical feasibility from safety specifications, allowing agents to improve capabilities through closed-loop execution feedback while converting probabilistic risk assessments into deterministic hard constraints for rigorous pre-action defense. Experimental results on the SafeAgentBench benchmark demonstrate that NEXUS achieves superior task success rates compared to existing methods. It effectively refuses unsafe instructions, exhibits robust defense against adversarial attacks, and progressively enhances planning efficiency through accumulated knowledge. This development represents a significant step forward in ensuring safe and reliable operation of AI-driven robots and embodied agents in real-world environments, bridging the critical divide between flexible language models and rigid safety requirements.
Wire timeline
NEXUS: Continual Learning of Symbolic Constraints for Safe and Robust Embodied Planning
Researchers have introduced NEXUS, a novel modular framework designed to address the safety gap between the probabilistic nature of Large Language Models (LLMs) and the strict determinism required in physical embodied intelligence. Unlike previous approaches that treat symbolic artifacts as static interfaces, NEXUS utilizes them for symbolic grounding and continuous knowledge evolution. The framework explicitly decouples physical feasibility from safety specifications, allowing agents to improve capabilities through closed-loop execution feedback while converting probabilistic risk assessments into deterministic hard constraints for rigorous pre-action defense. Experimental results on the SafeAgentBench benchmark demonstrate that NEXUS achieves superior task success rates compared to existing methods. It effectively refuses unsafe instructions, exhibits robust defense against adversarial attacks, and progressively enhances planning efficiency through accumulated knowledge. This development represents a significant step forward in ensuring safe and reliable operation of AI-driven robots and embodied agents in real-world environments, bridging the critical divide between flexible language models and rigid safety requirements.
cs.AI updates on arXiv.org