NEXUS: Continual Learning of Symbolic Constraints for Safe and Robust Embodied Planning

2026-05-10

AI Summary

This work addresses the fundamental tension between the probabilistic uncertainty inherent in large language models and the determinism and verifiable safety required in embodied intelligence. To reconcile this, the authors propose NEXUS, a modular framework that dynamically integrates symbolic constraints into continual learning, decoupling physical feasibility from safety specifications. By leveraging closed-loop feedback, NEXUS transforms risk assessments into deterministic hard constraints proactively—before action execution. The framework innovatively elevates symbolic artifacts from static interfaces to dynamic components capable of supporting symbol grounding and knowledge evolution. Experimental results on SafeAgentBench demonstrate that NEXUS significantly improves task success rates, effectively rejects unsafe instructions, resists adversarial attacks, and continuously enhances planning efficiency through accumulated knowledge.

Abstract

While Large Language Models (LLMs) have catalyzed progress in embodied intelligence, a fundamental gap between their inherent probabilistic uncertainty and the strict determinism and verifiable safety required in the physical world. To mitigate this gap, this paper introduces NEXUS, a modular framework designed for continual learning in embodied agents. Different from prior works that treat symbolic artifacts merely as static interfaces, NEXUS leverages them for symbolic grounding and knowledge evolution. The framework explicitly decouples physical feasibility from safety specifications: capability of agents is improved through closed-loop execution feedback, while probabilistic risk assessments are grounded into deterministic hard constraints to establish a rigorous pre-action defense. Experiments on SafeAgentBench demonstrate that NEXUS achieves superior task success rates while effectively refusing unsafe instructions, exhibiting robust defense against adversarial attacks, and progressively improving planning efficiency through knowledge accumulation.

Problem

• Continual Learning
• Symbolic Constraints
• Embodied Planning
• Safety
• LLMs

Innovation

• Continual Learning
• Symbolic Grounding
• Deterministic Safety Constraints
• Embodied Planning
• Modular Framework