Large language models let embodied agents turn natural-language instructions into plans, but those plans can be physically impossible or unsafe. Existing neuro-symbolic approaches tend to treat execution correction and safety enforcement as separate problems, so knowledge gained in one task is hard to reuse. The NEXUS framework instead keeps both a PDDL domain model and an LTL safety specification, updating them together as the agent interacts with the world.
When an action fails during execution, NEXUS repairs the corresponding action schema. When a hazard is detected, it converts the situation into a reusable LTL constraint by sampling semantic and translation variants hierarchically and keeping those that reach equivalence consensus. During planning, these constraints are enforced through LTL progression in symbolic search. By comparing constrained and unconstrained search outcomes, the system can distinguish physical unreachability from safety-induced refusal.
Experiments on the SafeAgentBench evaluation set show that NEXUS achieves the best overall trade-off between completing safe tasks and refusing unsafe ones, while keeping violation rates low even under jailbreak attacks. Ablations confirm that the PDDL domain refinement and the LTL constraint accumulation contribute complementary benefits. This summary reflects the claims of the single arXiv preprint; no independent verification is described in the abstract.