ChatHTN: Interleaving Approximate (LLM) and Symbolic HTN Planning

TL;DR

ChatHTN introduces a sound, hybrid HTN planner that interleaves symbolic task decomposition with context-rich prompts to ChatGPT for plausible decompositions when no method is available, thereby alleviating the knowledge-engineering bottleneck in HTN planning. The approach preserves correctness through verifier tasks that re-check task effects after ChatGPT-derived decompositions, ensuring generated plans satisfy input tasks. An open-source PyHop-based implementation demonstrates soundness across three domains, revealing when LLM assistance is needed and providing cost and variability insights. This work highlights a practical pathway to reliable hierarchical planning that leverages LLMs without sacrificing formal guarantees, with avenues for extending hierarchy depth and learning HTN methods from LLM-derived decompositions.

Abstract

We introduce ChatHTN, a Hierarchical Task Network (HTN) planner that combines symbolic HTN planning techniques with queries to ChatGPT to approximate solutions in the form of task decompositions. The resulting hierarchies interleave task decompositions generated by symbolic HTN planning with those generated by ChatGPT. Despite the approximate nature of the results generates by ChatGPT, ChatHTN is provably sound; any plan it generates correctly achieves the input tasks. We demonstrate this property with an open-source implementation of our system.

Figures (3)

• Figure 1: Initial hierarchical decomposition for a task from the logistics transportation domain. Boxed predicates are compound tasks and unboxed tasks are primitive tasks. ChatHTN stops on the planeTransport task because it lacks the knowledge to decompose it.
• Figure 2: ChatHTN prompts ChatGPT for a decomposition of the planeTransport task.
• Figure 3: ChatHTN continues hierarchical planning for the remaining task to generate a complete plan.