机器人系列报告一：具身智能：决定机器人泛化能力天花板的“大小脑”

Investment Rating - The report suggests a focus on the embodied intelligence industry, highlighting its potential as a ceiling for robot development, and recommends investors to pay attention to technological breakthroughs and commercialization progress in this sector [3]. Core Insights - Embodied intelligence is defined as an intelligent system based on physical bodies that perceives and acts through interaction with the environment, leading to intelligent behavior and adaptability [2][8]. - The development of embodied intelligence represents a convergence of robotics and artificial intelligence, with the potential to unlock significant market opportunities, projected to reach trillions in the future [2][3]. - The report identifies two main types of embodied large models: end-to-end models and layered embodied models, with the latter currently being more prevalent due to data limitations [2][3]. Summary by Sections What is Embodied Intelligence - Embodied intelligence emphasizes intelligent agents with physical bodies interacting with the environment to gain intelligence [8]. - The application of embodied intelligence in robotics can be divided into three stages: perception, reasoning, and execution [2][8]. Embodied Intelligence: Intersection of Robotics and AI - The generality of robots depends on the development of their generalization capabilities, which has evolved from simple automation to more complex, general-purpose robots [21][22]. - The report discusses the historical evolution of robotics from automation tools to intelligent agents that can directly impact the physical world [24]. Future of Humanoid Robots - The humanoid robot industry is expected to transition from specialized to general-purpose applications, with a significant focus on commercial viability in industrial manufacturing [2][3]. - The report highlights the anticipated investment boom in embodied intelligence, driven by advancements in AI and robotics [31][34]. Embodied Large Models Empowering Humanoid Robots - The report distinguishes between disembodied models and embodied large models, with the latter capable of generating motion postures linked to physical robots [37]. - It discusses the advantages and challenges of end-to-end models and layered models in the context of embodied intelligence [40][50]. Data Accumulation Methods for Robotics - The report outlines four primary methods for accumulating training data for robots: remote operation, AR, simulation, and video learning, emphasizing the current data scarcity in the robotics field [60][64][70].