Event

NEXUS: An Embodied Planning Framework Jointly Maintaining Action Models and Safety Constraints

1 reports · 1 sources

Tiehan Cui et al. propose NEXUS, a modular neuro-symbolic framework that jointly maintains a PDDL action domain model and LTL safety specifications: execution failures are used to repair action schemas, and detected hazards are converted into reusable LTL constraints via hierarchical semantics with translation sampling and equivalence consensus. The authors report that on the SafeAgentBench evaluation set, NEXUS achieves the best overall trade-off between completing safe tasks and rejecting unsafe tasks, and maintains a lower violation rate under jailbreak attacks; ablation experiments show that PDDL domain refinement and LTL constraint accumulation have complementary effects.