REFLEX: Metacognitive Reasoning for Reflective Zero-Shot Robotic Planning with Large Language Models

TL;DR

REFLEX is presented, a framework that integrates metacognitive learning into LLM-powered multi-robot collaboration and shows results that support the hypothesis that metacognitive learning can foster creativity in robotic planning.

Abstract

While large language models (LLMs) have shown great potential across various domains, their applications in robotics remain largely limited to static prompt-based behaviors and still face challenges in complex tasks under zero-shot or few-shot settings. Inspired by human metacognitive learning and creative problem-solving, we address this limitation by exploring a fundamental question: Can LLMs be empowered with metacognitive capabilities to reason, reflect, and create, thereby enhancing their ability to perform robotic tasks with minimal demonstrations? In this paper, we present REFLEX, a framework that integrates metacognitive learning into LLM-powered multi-robot collaboration. The system equips the LLM-powered robotic agents with a skill decomposition and self-reflection mechanism that identifies modular skills from prior tasks, reflects on failures in unseen task scenarios, and synthesizes effective new solutions. We propose a more challenging robotic benchmark task and evaluate our framework on the existing benchmark and the novel task. Experimental results show that our metacognitive learning framework significantly outperforms existing baselines. Moreover, we observe that our framework can generate solutions that differ from the ground truth yet still successfully complete the tasks. These findings support our hypothesis that metacognitive learning can foster creativity in robotic planning.

Figures (4)

• Figure 1: Overview of the proposed REFLEX framework. The framework consists of three interconnected components: (1) modular skill set construction, where prior successful task exemplars are decomposed and clustered into reusable modular manipulation skills; (2) metacognitive inference, where relevant skills and associated exemplars are selectively retrieved and composed to synthesize motion plans for unseen tasks; and (3) structured self-reflection, where validation feedback updates the metacognition-informed input to diagnose missing or misapplied skills and regenerate improved plans. This closed-loop reasoning mechanism enables reliable recovery and adaptive alternative plan generation under zero-shot multi-robot collaboration settings.
• Figure 2: We create a new benchmark robotic task: Install Drywall. Together with the three most challenging tasks in RoCoBench mandi2024roco, we test and compare our metacognitive learning framework on the four tasks with the baseline methods.
• Figure 3: A case study of generating a creative and valid solution by our metacognitive learning framework REFLEX.
• Figure 4: Demonstration of our metacognitive inference and self-reflection components. In most cases, the metacognitive inference component works well for generating valid subtask plans; if not, the self-reflection component can serve as a fix.