HiCRISP: An LLM-Based Hierarchical Closed-Loop Robotic Intelligent Self-Correction Planner

Abstract

The integration of Large Language Models (LLMs) into robotics has revolutionized human-robot interactions and autonomous task planning. However, these systems are often unable to self-correct during the task execution, which hinders their adaptability in dynamic real-world environments. To address this issue, we present an LLM-based Hierarchical Closed-loop Robotic Intelligent Self-correction Planner (Hi-CRISP), an innovative framework that enables robots to correct errors within individual steps during the task execution. HiCRISP actively monitors and adapts the task execution process, addressing both high-level planning and low-level action errors. Extensive benchmark experiments, encompassing virtual and real-world scenarios, showcase HiCRISP's exceptional performance, positioning it as a promising solution for robotic task planning with LLMs.