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

Researchers introduce NEXUS, a framework enabling embodied AI agents to learn symbolic constraints for safer decision-making in physical environments. The system addresses the gap between probabilistic language models and the deterministic safety requirements of robotics by decoupling physical feasibility from safety specifications, achieving improved task success while refusing unsafe instructions.

NEXUS represents a meaningful advancement in embodied AI safety, tackling a critical challenge that has limited real-world deployment of language model-driven robotics. Large language models excel at reasoning but generate probabilistic outputs fundamentally misaligned with physical world requirements where safety failures carry tangible consequences. The framework's core innovation lies in treating symbolic artifacts as active learning mechanisms rather than static interfaces, enabling agents to evolve their knowledge through closed-loop execution feedback while maintaining rigorous safety guardrails.

This work emerges from growing recognition that scaling language models alone cannot solve embodied intelligence. Prior approaches either ignored safety specifications entirely or treated them as afterthoughts, creating systems that performed tasks efficiently but lacked trustworthy refusal mechanisms for dangerous instructions. NEXUS explicitly decouples these concerns, improving physical capability through experience while grounding probabilistic risk assessments into hard constraints that function as pre-action defense layers.

The experimental validation on SafeAgentBench demonstrates practical utility beyond theoretical elegance. Superior task success combined with robust adversarial defense and progressive planning efficiency gains suggests the framework could enable safer deployment of autonomous systems in manufacturing, healthcare, and domestic robotics. The ability to refuse unsafe instructions while maintaining performance represents a significant step toward trustworthy automation.

Key development challenges ahead include scaling the approach to more complex multi-agent scenarios, handling real-world sensor uncertainty, and establishing formal verification methods for symbolic constraints. The framework's modular design suggests compatibility with emerging robotics standards, potentially influencing how future embodied AI systems balance capability with safety.

• →NEXUS decouples physical feasibility from safety specifications, enabling robots to learn capabilities while maintaining strict safety constraints.
• →The framework treats symbolic knowledge as dynamically evolving rather than static, improving planning efficiency through cumulative learning.
• →Experimental results show superior task success rates with robust refusal of unsafe instructions and adversarial attack defense.
• →The approach addresses fundamental misalignment between probabilistic language models and deterministic physical world requirements.
• →Modular design suggests potential integration with broader robotics ecosystems and autonomous systems development.