DUPLEX: Reliable LLM Robotic Planning

🔑 Enhanced Key Takeaways

• •DUPLEX utilizes a neuro-symbolic feedback loop where the PDDL solver's error messages are parsed back into natural language prompts to guide the 'Slow System' in iterative plan refinement.
• •The architecture specifically addresses the 'hallucination of impossible actions' common in end-to-end LLM planners by enforcing strict adherence to a predefined PDDL domain schema during the extraction phase.
• •Empirical evaluations indicate that DUPLEX significantly reduces the computational overhead compared to monolithic LLM approaches by offloading complex reasoning to symbolic solvers, which are inherently more efficient for combinatorial search.

🛠️ Technical Deep Dive

• •Architecture: Dual-pathway system consisting of a 'Fast Path' (distilled LLM, e.g., Llama-3-8B) for rapid PDDL translation and a 'Slow Path' (high-capacity LLM, e.g., GPT-4o or Claude 3.5 Sonnet) for diagnostic reasoning.
• •Integration: Uses a standard PDDL 2.1 parser to validate generated plans against domain constraints before execution.
• •Feedback Mechanism: Implements a 'Diagnostic-to-Prompt' bridge that translates solver-specific error codes (e.g., 'precondition not met') into semantic feedback for the Slow System.
• •Constraint Satisfaction: Employs a schema-guided extraction layer that constrains the LLM's output space to the specific predicates and objects defined in the PDDL domain file.

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline