NEXUS: LLM Framework for Safe Embodied Planning
NEXUS, an innovative modular framework, facilitates ongoing learning in embodied agents by separating physical feasibility from safety requirements. In contrast to earlier methods that view symbolic artifacts as fixed interfaces, NEXUS employs them for symbolic grounding and knowledge advancement. This framework enhances agent performance via closed-loop execution feedback and translates probabilistic risk evaluations into firm constraints for preemptive safety measures. Tests conducted on SafeAgentBench reveal that NEXUS not only attains higher task success rates but also adeptly declines unsafe directives, bridging the divide between the probabilistic uncertainty of LLMs and the stringent safety needed for tasks in the physical realm.
Key facts
- NEXUS is a modular framework for continual learning in embodied agents.
- It decouples physical feasibility from safety specifications.
- Capability is improved through closed-loop execution feedback.
- Probabilistic risk assessments are grounded into deterministic hard constraints.
- Experiments were conducted on SafeAgentBench.
- NEXUS achieves superior task success rates.
- It effectively refuses unsafe instructions.
- The framework addresses the gap between LLM uncertainty and physical world safety.
Entities
—