从语言到动作:大型语言模型赋能的机器人操纵范式综述OA

Traditional robot control in unstructured environments is often hindered by programming complexity,limited generalization,and heavy reliance on domain experts.This paper provides a systematic review of robot control empowered by large language models(LLMs),summarizing recent progress and major technical pathways.This paper first identifies the"three gaps"in instruction,data,and knowledge that constrain traditional paradigms,and discusses how LLMs enable the construction of a unified cognitive core for robotics through semantic understanding,commonsense reasoning,and code generation.The main contribution is an in-depth synthesis and comparison of three mainstream language-to-action paradigms.Unlike prior surveys that are commonly organized by functional modules(language understanding,visual per-ception,and motion control),this paper emphasizes the differences among paradigms in physical-world grounding and execution-feedback closed-loop mechanisms.Specifically,this paper reviews:(1)high-level planners that output symbolic action sequences and offer reliability and interpretability;(2)code generators that produce executable programs and improve flexibility for complex logic and long-horizon tasks;and(3)end-to-end policies that directly map perception to actions and show strong potential for building generalist physical agents.Meanwhile,this paper systematically reviews and incorporates recent progress on end-to-end vision-language-action policies,thereby improving the coverage and time-liness of the survey.This paper further analyzes key challenges,including grounding,safety and reliability,dynamic feed-back and closed-loop control,and real-time performance versus computational efficiency.This paper concludes that LLMs are reshaping robot control toward an intent-driven paradigm centered on natural language,and outlines future direc-tions in multimodal embodied intelligence,proactive human-robot collaboration,and efficient edge deployment.