Language-Augmented Symbolic Planner for Open-World Task Planning

TL;DR

The paper addresses open-world robotic task planning under incomplete domain knowledge by introducing LASP, which combines symbolic planning with pre-trained LLMs. LASP uses observation-driven error diagnosis to iteratively refine action preconditions, properties, and objects, and replans to accomplish goals in uncertain environments. It defines a formal planning framework with incomplete knowledge, and demonstrates through kitchen-domain experiments that LASP outperforms several language-driven baselines, particularly on long-horizon tasks. The results show LASP preserves the interpretability and verifiability of symbolic planning while leveraging LLMs to fill knowledge gaps, with practical implications for robust autonomous agents in open-world settings.

Abstract

Enabling robotic agents to perform complex long-horizon tasks has been a long-standing goal in robotics and artificial intelligence (AI). Despite the potential shown by large language models (LLMs), their planning capabilities remain limited to short-horizon tasks and they are unable to replace the symbolic planning approach. Symbolic planners, on the other hand, may encounter execution errors due to their common assumption of complete domain knowledge which is hard to manually prepare for an open-world setting. In this paper, we introduce a Language-Augmented Symbolic Planner (LASP) that integrates pre-trained LLMs to enable conventional symbolic planners to operate in an open-world environment where only incomplete knowledge of action preconditions, objects, and properties is initially available. In case of execution errors, LASP can utilize the LLM to diagnose the cause of the error based on the observation and interact with the environment to incrementally build up its knowledge base necessary for accomplishing the given tasks. Experiments demonstrate that LASP is proficient in solving planning problems in the open-world setting, performing well even in situations where there are multiple gaps in the knowledge.

Figures (5)

• Figure 1: An illustrative example of the challenge in open-world task planning is that the knowledge (i.e., domain and problem files shown in (a) and (b)) of the symbolic planner can be incomplete. This causes the execution failure of the planned action as depicted in (c).
• Figure 2: An overview of our proposed LASP, a task planning framework in the open world. The symbolic planner is responsible for finding a plan to accomplish the given task, and subsequently, the robot executes this plan. Given an initial, incorrect plan, the robot will encounter some errors during action execution. Our proposed method LASP will invoke the LLM to recursively refine the task planning problem by supplementing action knowledge to the planner. Moreover, if the planner is unable to find a plan afterward, the LLM can augment the planner's task-specific knowledge with necessary objects to assist the planner in finding an error-free solution.
• Figure 3: Qualitative result of LASP for the task of heating milk. LASP is able to refine incomplete knowledge within the planning task and find a plan to accomplish the task. The snapshots of the execution actions are taken from our simulation platform built on PyRep james2019pyrep. Comments are indicated in gray text and start with a semicolon.
• Figure 4: Two code examples used in ProgPrompt.
• Figure 5: A code example generated by ProgPrompt.