Aligning LLM+PDDL Symbolic Plans with Human Objective Specifications through Evolutionary Algorithm Guidance

TL;DR

This work tackles aligning LLM-generated PDDL plan specifications with human objectives when NL translations produce imperfect symbolic constraints. It introduces a neurosymbolic pipeline that uses a genetic algorithm to evolve plan specifications, guided by an LSTM-based specification adherence model and a symbolic planner (OPTIC) to generate state-trajectories. Empirical results in a naval disaster scenario and a satellite-domain validation show that GA-guided refinement yields more adherent plans than LLM-only translations, improving valid-plan rates from roughly 32–6% to about 48–34%. The approach enables non-experts to specify goals in natural language while systematically correcting mistranslations, with potential for application to dynamic, safety-critical planning tasks.

Abstract

Automated planning using a symbolic planning language, such as PDDL, is a general approach to producing optimal plans to achieve a stated goal. However, creating suitable machine understandable descriptions of the planning domain, problem, and goal requires expertise in the planning language, limiting the utility of these tools for non-expert humans. Recent efforts have explored utilizing a symbolic planner in conjunction with a large language model to generate plans from natural language descriptions given by a non-expert human (LLM+PDDL). Our approach performs initial translation of goal specifications to a set of PDDL goal constraints using an LLM; such translations often result in imprecise symbolic specifications, which are difficult to validate directly. We account for this using an evolutionary approach to generate a population of symbolic goal specifications with slight differences from the initial translation, and utilize a trained LSTM-based validation model to assess whether each induced plan in the population adheres to the natural language specifications. We evaluate our approach on a collection of prototypical specifications in a notional naval disaster recovery task, and demonstrate that our evolutionary approach improve adherence of generated plans to natural language specifications when compared to plans generated using only LLM translations. The code for our method can be found at https://github.com/owenonline/PlanCritic.

Figures (5)

• Figure 1: Diagram of the proposed system. Natural language feedback to an initial plan provided by the user are translated to an initial plan specification. The genetic algorithm searches locally over the initial plan specification, producing a population of candidate specifications. The specification adherence model quantifies the extent to which plans adhere to the user's natural language feedback, and is used to guide the genetic algorithm search.
• Figure 2: Example interface demonstrating needed aspects for user-system interaction.
• Figure 3: Grammar used to define feedback as PDDL constraints.
• Figure 4: Naval disaster response scenario, showing two waterways separating assets from a target ship to salvage. Waterways are blocked by underwater debris (black), or floating debris (red), that must be removed prior to traversing the waterway. Underwater debris must be discovered by the scout asset before removal.
• Figure 5: Overall GA success rate based on the non-convergence rate of the GA (y-axis) and failures due to false positives of the adherence model (x-axis).