物理人工智能，第1部分：基础知识

Physical AI, part 1: The basics Key takeaways •Physical AI marks the next major phase of AI commercialization, extending intelligence beyond software and into machines thatcan see, decide and act in the real world - unlocking multi-trillion-dollar opportunities across robotics, autonomous vehicles anddrones. •Progress is accelerating due to converging advances in models, data, compute and simulation, with multimodal foundationmodels, synthetic data and world models enabling robots to learn, interact and train more safely and efficiently. •With these advancements, we're seeing robotics shift from rules‑based systems to data‑driven, end‑to‑end AI, supported byopen‑source innovation and improved observability paving the way for adaptable, general‑purpose robots and more autonomousindustrial systems. •This is the first publication in a three-part series that aims to explore what happens when AI leaves the chat, and shows up inthe real world. What is physical AI?Physical AI, or embodied AI, refers to the shift from systems that exist only in digital environments to machines that can sense, reason, and act in the physical world–such as humanoid robots, autonomous vehicles and drones. This evolution marks the nextmajor phase of AI commercialization, pushing intelligence beyond software and digital output, like chatbots and contentgeneration, and into machines capable of manipulating their surroundings and delivering real-world productivity gains (Exhibit 1).New to these terms? ReadAI Dictionary, part 1: The basicsfor quick primers on core concepts. This transition, however, didn’t happen overnight. As BofA Global Research notes, the rise of physical AI stems from cumulativetechnological progress, rather than a single breakthrough. Physical AI systems combine sensors and computer vision with AImodels that learn through interaction. Rather than relying on explicit programming for every action, these systems learn andimprove through trial, feedback and repeated experience. For instance, instead of hard-coding a robot to pick up an object,physical AI enables it to experiment, adjust and refine its approach until it masters the task–resulting in far greater adaptabilityand autonomy. Early applications focus on automating repetitive or hazardous tasks in areas such as warehouse operations, manufacturing andlogistics. As self-learning capabilities improve, physical AI is expected to evolve from task-based automation toward greaterlevels of autonomy–reshaping mobility, fulfilment, and industrial operations in the process. And people are taking notice. Coverage, interest, and public awareness of physical AI have been growing fast across mainstreambusiness and financial media, social platforms, and investors and corporates as they focus on robotics, autonomy, and AImonetization beyond software. BofA Global Research notes that news mentions and market commentary around physical AI continue to rise, with January 2026 generating the highest number of citations in the past two years and almost nine times asmany citations as January 2024 (Exhibit 2). Meanwhile, social engagement on X is also trending higher (Exhibit 3). Number of physical AI citations across trade publications and newssources Technology enablement: Scaling beyond LLMsRapid cost reductions and performance improvements across AI models, chips, sensors, batteries, and manufacturing capacity are enabling the development of machines that can see, decide, and act in real time. Together, these advances are unlockingwhat could be multi-trillion-dollar opportunities across robotics, autonomous vehicles, drones and the broader AI infrastructure. Exhibit4:A combination of better models, data and compute is enabling more sophisticated AI applicationsInfographic illustrating how AI is scaling beyond large language models (LLMs) According to BofA Global Research, advancing physical AI will require fundamentally new model architectures, training methods,and compute strategies–approaches that go beyond the techniques that previously scaled LLMs. However, several convergingtrends are now accelerating the types of tasks AI can perform, helping explain why progress is spilling over from digital cognitioninto the physical AI world (Exhibit 4). Trends enabling the rise of physical AI include:1 •Compute and data: Advances in model architecture demand massive amounts of compute and data. Despite limitations,scaling laws continue to hold: models improve as data and compute increase. To supplement limited real-worldinteraction data, synthetic data generation is expanding rapidly, providing richer training signals for physical tasks. •Agentic AI: New AI systems are becoming more autonomous, able to gather information, conduct research, plan andsimulate outcomes on their own. These capabilities allow models to perform“inference-time compute”as it is required,often at the edge, enabling more adaptive and context-aware decision-making. For deeper reading, we covered the la