Neuro-Symbolic Task Planning and Replanning using Large Language Models
Keywords: Robotics, Task Planning, Large Language Models
TL;DR: A neuro-symbolic task planning and replanning framework that integrates LLMs and symbolic planners to improve efficiency, accuracy, and failure recovery in complex robotic tasks.
Abstract: In robotic task planning, symbolic planners are robust but struggle with long, complex tasks because their search space grows exponentially. By contrast, approaches based on LLMs reason faster and incorporate commonsense knowledge but show lower success rates and lack failure recovery.
We present a novel neuro-symbolic task planning framework with subgoal decomposition to overcome the drawbacks of symbolic planners (slow speed) and LLM-based methods (low accuracy). It breaks down complex tasks into subgoals using an LLM, then selects either a symbolic planner or an MCTS-based LLM planner to handle each subgoal according to its complexity.
Furthermore, we propose a neuro-symbolic task replanning algorithm for task planning failure recovery. During task planning and low-level code generation, LLM performs as syntax and semantic checkers to ensure validity and facilitate replanning if necessary.
We demonstrate that both task planning and replanning improve high success across diverse PDDL domains, as well as in real and simulated robotics environments.
More details are available at http://graphics.ewha.ac.kr/LLMTAMP/
Submission Number: 8
Loading