中金 | 人机系列04：具身智能大脑的进化之路

Core Viewpoint - The report emphasizes that the field of embodied intelligence is transitioning from "route differentiation" to "fusion landing," with data-driven and heterogeneous training becoming central to achieving general intelligence [2]. Group 1: Embodied Intelligence Algorithms - The evolution of robot algorithms has shifted from model-driven to data-driven approaches, with hierarchical control as a foundational paradigm and VLA (Vision-Language-Action) models enhancing generalization and interaction capabilities [5][7]. - The current mainstream paths in robot algorithms include hierarchical architecture, VLA, and world models, each differing in theoretical logic, engineering implementation, and industrial application [9][11]. - The introduction of large model structures, such as Transformers, has established a unified algorithmic foundation, facilitating cross-task learning and creating a closed loop of perception, cognition, and control [8][9]. Group 2: Data in Embodied Intelligence - Data acquisition for humanoid robots has evolved into three main paths: real machine acquisition, video learning, and simulation generation, forming a complementary ecosystem [16][18]. - Real machine acquisition emphasizes high-value, high-cost feedback through teleoperation, while video learning leverages low-cost, high-diversity visual data to enhance training [20][22]. - Simulation-generated data is becoming a significant source for large-scale training, with advancements in high-fidelity physics engines and digital twin environments facilitating the Sim2Real transition [23][24]. Group 3: Hot Topics in Embodied Intelligence - The Scaling Law phenomenon in embodied intelligence indicates that as model size, data, and computing power increase, robots significantly improve in cognition and behavior, leading to breakthroughs in generalization and task capabilities [27][28]. - The lack of standardized benchmarks for evaluating embodied intelligence poses challenges, with recent efforts like the BEHAVIOR-1K benchmark aiming to establish a comprehensive evaluation framework [29][30]. - Physical AI, which integrates physical knowledge with AI models, is emerging as a foundational exploration direction, enhancing robots' understanding of physical rules and causal reasoning [35][37]. Group 4: Industry Landscape - The humanoid robot software ecosystem comprises foundational models, data science, simulation software, and evaluation systems, with major tech companies and startups collaborating to build this ecosystem [45][46]. - Key players in the industry include tech giants like Google, Meta, and NVIDIA, alongside humanoid robot startups that focus on AI model development and data acquisition systems [45][46].