From Conversation to Action: Opportunities and Challenges of Large Language Models as the Brain of Humanoid Robots

Abstract

Humanoid robots, as the core carriers of embodied intelligence, are moving from specialized scenarios to general-purpose applications. The incorporation of large language models (LLMs) endows them with cognitive and decision-making capabilities similar to those of a human "brain," becoming a key enabler for technological breakthroughs. Based on the practical and technological achievements of the global humanoid robot industry from 2024 to 2025, this paper systematically analyzes the technical value and practical bottlenecks of LLMs as robot brains. The study finds that, through semantic parsing, task planning, and generalized learning capabilities, LLMs increase the success rate of task decomposition in complex scenarios to over 95%, reducing deployment costs by 60%. However, they also face core challenges such as data shortages, insufficient real-time performance, and poor architectural adaptability. Drawing on practical examples from Google's RT series, Tesla's Optimus, and domestic companies such as iFlytek and UBTECH, this paper proposes a three-step breakthrough path: "data accumulation - architecture optimization - ecosystem collaboration." Research indicates that the deep integration of LLM and robotics will accelerate the scale-up of the industry. The Chinese market is expected to reach 75 billion yuan in 2029, accounting for 32.7% of the global market.