Grounding Generative Planners in Verifiable Logic: A Hybrid Architecture for Trustworthy Embodied AI

Large Language Models (LLMs) show promise as planners for embodied AI, but their stochastic nature lacks formal reasoning, preventing strict safety guarantees for physical deployment. Current approaches often rely on unreliable LLMs for safety checks or simply reject unsafe plans without offering repairs. We introduce the Verifiable Iterative Refinement Framework (VIRF), a neuro-symbolic architecture that shifts the paradigm from passive safety gatekeeping to active collaboration. Our core contribution is a tutor-apprentice dialogue where a deterministic Logic Tutor, grounded in a formal safety ontology, provides causal and pedagogical feedback to an LLM planner. This enables intelligent plan repairs rather than mere avoidance. We also introduce a scalable knowledge acquisition pipeline that synthesizes safety knowledge bases from real-world documents, correcting blind spots in existing benchmarks. In challenging home safety tasks, VIRF achieves a perfect 0 percent Hazardous Action Rate (HAR) and a 77.3 percent Goal-Condition Rate (GCR), which is the highest among all baselines. It is highly efficient, requiring only 1.1 correction iterations on average. VIRF demonstrates a principled pathway toward building fundamentally trustworthy and verifiably safe embodied agents.

💡 Research Summary

The paper tackles the critical safety gap that arises when large language models (LLMs) are used as planners for embodied agents. While LLMs excel at generating flexible, high‑level plans, their stochastic nature and lack of formal reasoning prevent the provision of deterministic safety guarantees required for real‑world deployment. Existing approaches either rely on self‑reflection mechanisms (e.g., chain‑of‑thought, constitutional prompting) or on external tool‑use, but both ultimately depend on the same unreliable LLM to supervise its own output, leading to a self‑referential loop with no provable safety.

To break this loop, the authors introduce the Verifiable Iterative Refinement Framework (VIRF), a neuro‑symbolic architecture that couples a generative LLM planner (the “Apprentice”) with a deterministic Logic Tutor grounded in a formal safety ontology expressed in OWL‑2. The Tutor acts as a System‑2 verifier: it receives the LLM‑generated plan, translates it into a symbolic representation, and checks it against the ontology using description‑logic inference. If the plan violates any safety constraint, the Tutor does not merely reject it; instead, it produces a structured diagnostic report that includes a full proof trace, explicitly stating why the action is hazardous (e.g., “the microwave is on and the pot is metallic, which can cause a spark”). This causal, pedagogical feedback teaches the LLM the underlying principles, enabling it to repair the plan intelligently rather than abandoning the task.

A major bottleneck for any logic‑based system is the construction of a comprehensive, verifiable knowledge base. VIRF addresses this with the Traceable Axiom Synthesis (TAS) pipeline. TAS automatically extracts safety statements from real‑world documents (regulations, manuals, product datasheets), retrieves relevant text snippets, uses an LLM to draft candidate axioms in formal syntax while citing the source sentences, and finally hands the draft to a human expert for semantic and logical validation. This human‑in‑the‑loop step guarantees that each axiom faithfully captures the nuanced intent of the source material, while the LLM handles the heavy lifting of discovery and formalization. The authors report synthesizing 92 verified axioms in two days, dramatically expanding the coverage of their safety ontology.

On the perception side, VIRF introduces a VLM‑Cascade perception pipeline that builds a Rich Semantic Scene Graph (RSSG). The three‑stage process first extracts objects and coarse relations using a vision‑language model, then enriches each node with fine‑grained attributes critical for safety (temperature, conductivity, material state), and finally maps these attributes to the ontology. This detailed scene graph allows the Logic Tutor to reason about subtle hazards that generic scene‑graph generators miss.

The experimental evaluation uses a newly created suite of home‑environment safety tasks (12 scenarios such as “microwave with metal pot”, “chemical spill near children”, etc.). VIRF achieves a Hazardous Action Rate (HAR) of 0 %—no unsafe actions are executed—while attaining a Goal‑Condition Rate (GCR) of 77.3 %, the highest among all baselines. The average number of correction iterations per episode is only 1.1, indicating that the Tutor’s diagnostic reports dramatically reduce the search space for the LLM. In contrast, self‑refinement methods and tool‑augmented agents either retain a non‑zero HAR or require many more iterations to converge.

A further contribution is the systematic analysis of blind spots in existing safety benchmarks. By comparing the coverage of the automatically built ontology against standard datasets, the authors show that many chemical, food, and electrical hazards are missing, and that their TAS‑generated axioms fill over 30 % of these gaps. This highlights the importance of scalable, traceable knowledge acquisition for trustworthy embodied AI.

In summary, VIRF demonstrates a principled pathway to trustworthy embodied agents by (1) anchoring stochastic LLM planning in a deterministic, formally verified logic core, (2) providing causal, pedagogical feedback that enables plan repair rather than rejection, and (3) automating the creation of a rich, verifiable safety ontology from real‑world texts. The work bridges the gap between generative flexibility and formal safety, offering a scalable blueprint for future deployments of AI agents in safety‑critical physical environments.