Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PDDLEGO: Iterative Planning in Textual Environments

Li Zhang, Peter Jansen, Tianyi Zhang, Peter Clark, Chris Callison-Burch, Niket Tandon

TL;DR

The paper tackles planning under partial observability in textual environments, where end-to-end LLM planning struggles due to incomplete knowledge. It introduces PDDLEGO, a neurosymbolic framework that iteratively builds a PDDL representation during exploration by employing two LL-based modalities: PDDL-gen to generate a full problem file and PDDL-edit to apply constrained edits to the current PF, guided by sub-goals when the end-goal is unattainable. Empirically, PDDLEGO improves planning efficiency and success rate across two text-game benchmarks (Coin Collector and Cooking World) compared with end-to-end action generation, achieving a 43% efficiency gain on Coin Collector and up to 98% success in Cooking World easy, with substantial robustness in harder variants. The approach also improves interpretability and correctability by constraining the planning task to a deterministically solvable PF, albeit at the cost of slower PDDL generation and a requirement for domain-file annotations and sub-goal structures.

Abstract

Planning in textual environments have been shown to be a long-standing challenge even for current models. A recent, promising line of work uses LLMs to generate a formal representation of the environment that can be solved by a symbolic planner. However, existing methods rely on a fully-observed environment where all entity states are initially known, so a one-off representation can be constructed, leading to a complete plan. In contrast, we tackle partially-observed environments where there is initially no sufficient information to plan for the end-goal. We propose PDDLEGO that iteratively construct a planning representation that can lead to a partial plan for a given sub-goal. By accomplishing the sub-goal, more information is acquired to augment the representation, eventually achieving the end-goal. We show that plans produced by few-shot PDDLEGO are 43% more efficient than generating plans end-to-end on the Coin Collector simulation, with strong performance (98%) on the more complex Cooking World simulation where end-to-end LLMs fail to generate coherent plans (4%).

PDDLEGO: Iterative Planning in Textual Environments

TL;DR

The paper tackles planning under partial observability in textual environments, where end-to-end LLM planning struggles due to incomplete knowledge. It introduces PDDLEGO, a neurosymbolic framework that iteratively builds a PDDL representation during exploration by employing two LL-based modalities: PDDL-gen to generate a full problem file and PDDL-edit to apply constrained edits to the current PF, guided by sub-goals when the end-goal is unattainable. Empirically, PDDLEGO improves planning efficiency and success rate across two text-game benchmarks (Coin Collector and Cooking World) compared with end-to-end action generation, achieving a 43% efficiency gain on Coin Collector and up to 98% success in Cooking World easy, with substantial robustness in harder variants. The approach also improves interpretability and correctability by constraining the planning task to a deterministically solvable PF, albeit at the cost of slower PDDL generation and a requirement for domain-file annotations and sub-goal structures.

Abstract

Planning in textual environments have been shown to be a long-standing challenge even for current models. A recent, promising line of work uses LLMs to generate a formal representation of the environment that can be solved by a symbolic planner. However, existing methods rely on a fully-observed environment where all entity states are initially known, so a one-off representation can be constructed, leading to a complete plan. In contrast, we tackle partially-observed environments where there is initially no sufficient information to plan for the end-goal. We propose PDDLEGO that iteratively construct a planning representation that can lead to a partial plan for a given sub-goal. By accomplishing the sub-goal, more information is acquired to augment the representation, eventually achieving the end-goal. We show that plans produced by few-shot PDDLEGO are 43% more efficient than generating plans end-to-end on the Coin Collector simulation, with strong performance (98%) on the more complex Cooking World simulation where end-to-end LLMs fail to generate coherent plans (4%).
Paper Structure (11 sections, 6 equations, 8 figures, 2 tables)

This paper contains 11 sections, 6 equations, 8 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Environments
  4. Evaluation
  5. Conclusion
  6. Formulation of PDDL
  7. Annotated Domain Files and Prompts
  8. Hyperparameters
  9. Example Trajectories
  10. Generated Problem Files
  11. Results on the Development Set

Figures (8)

  • Figure 1: A fully-observed environment like BlocksWorld (upper, to rearrange objects from and to a given configuration) can be tackled by generating a PDDL problem file, while a partially observed one like Coin Collector (lower, to look for an object in an unknown location) cannot until sufficient exploration.
  • Figure 2: The pipeline of pddlego. A PDDL problem file is iteratively built during exploration.
  • Figure 3: On Coin Collector, the mean and standard deviation of number of steps to success (less is better) for each development example, each over 5 trials with different random seeds of gpt-4-1106-preview, comparing Action-gen and PDDL-edit. The error bar represents the sample standard deviation. On example 0 and 6, PDDL-edit fails and thus not shown.
  • Figure 4: A PDDL solver produces a plan based on a minimal domain file and problem file. Previous work assumes the domain file as given, while we predict the action definitions in the domain file.
  • Figure 5: Annotated domain file for Coin Collector.
  • ...and 3 more figures