大摩重磅机器人年鉴（二）：机器人"逃离工厂"，训练重点从“大脑”转向“身体”，边缘算力有望爆发

Core Insights - The article highlights a significant shift in the robotics industry, driven by artificial intelligence, moving from traditional factory settings to broader applications in homes, cities, and even space. This transition emphasizes the need for physical manipulation capabilities over cognitive abilities, which is expected to lead to a surge in demand for edge computing [1][2]. Group 1: Key Transformations in Robotics - The report identifies two major transformations in the global robotics industry: the escape of robots from structured factory environments to unstructured settings like homes and cities, and a shift in training focus from AI "brains" (general models) to "bodies" (physical action control) [1][3]. - Traditional industrial robots were limited to repetitive tasks in controlled environments, while AI-enabled robots are now capable of navigating complex real-world scenarios, such as autonomous vehicles in traffic and service robots in homes [3]. Group 2: Challenges in Physical Interaction - The article uses the example of "grabbing a bottle from the fridge" to illustrate the complexities of physical interactions, which involve multiple variables such as precise finger positioning, body balance, grip strength, and environmental factors [6]. - Robots must develop real-time perception, dynamic decision-making, and fine motor control capabilities, moving beyond reliance on pre-programmed instructions [7]. Group 3: Data Collection for Training - Unlike large language models that primarily use text and image data, robotic models require extensive real-world physical operation data, making data collection and model training more complex and costly [9]. - Major tech companies like Tesla, NVIDIA, and Google are employing three main methods to gather training data: teleoperation, simulation, and video learning [11]. Group 4: Edge Computing Demand - As robots transition from factories, the latency issues of centralized cloud computing become apparent, making edge computing a necessity. The report outlines two trends in edge computing: the proliferation of specialized edge chips and distributed inference networks [19][22]. - NVIDIA's Jetson Thor is highlighted as a representative edge real-time inference device, priced around $3,500, which has been adopted by companies like Boston Dynamics and Amazon Robotics for its high computational power at low energy consumption [19]. - Tesla's concept of "robots as computing nodes" suggests that deploying 100 million robots with 2,500 TFLOPS of computing power could provide a total of 125,000 ExaFLOPS, equivalent to 7 million NVIDIA B200 GPUs, enhancing overall efficiency through collaboration among robots [22]. Group 5: Future Projections - Morgan Stanley predicts that by 2030, global demand for edge computing in robotics will significantly increase, with various forms of robots contributing to substantial computational needs. By 2050, it is estimated that 1.4 billion robots will be sold globally, driving edge AI computing demand to the equivalent of millions of B200 chips [25].