物理人工智能

Executive Summary While the world has focusedits attention for the last three yearson generativeartificial intelligence, chatbots, andnewmodelreleases coming fromfrontierAIlabs, aquieter revolution is taking placethat many believerepresents the nextstage in AIdevelopment: the arrival ofPhysical AI.Likethe iPhone’s introduction in 2007,AlexNet’s victory in the 2012 ImageNet competition, and ChatGPT’s release in 2022,analystsand industry representativesbelieveasimilarbreakthrough is imminent. Physical AI “lets autonomous systems like robots, self-driving cars, and smart spacesperceive, understand, and perform complex actions in the real (physical) world.”1NVIDIAhas declared “in the near future, everything that moves, or that monitors thingsthat move, will be autonomous robotic systems.”2OpenAI reportedly re-opened itsrobotics divisionin early 2025to capitalize on the convergenceof AI and robotics,while startupsfromShanghai toSilicon Valley building the “brains” of robots areraising hundreds of millions of dollars.3Electric vehicle makersTeslaandXPengareracing to develophumanoid robotsof their own.4 Meanwhile Amazon, which reportshavingone millionrobots in operation today, believes “Physical AI is about to changeeverything for robotics [including] autonomy, manipulation, sortation, and computervision.”5Adding to this enthusiasm,analysts at Morgan Stanleyassertthe marketforhumanoid robotswillgrow fromtens of millions of dollarstoday toreach $5trillion by2050.6 Yet theconvergence of AI and robotics is so newthatthefield lacksa shared name, tosay nothing of a maturetechnology stack.Some companies callthis convergence“embodied AI” whileothers prefer “physical AI,”“embodied machine intelligence,” or“generative physical AI.”7It is notat allclear if the hype around AI progress cantranslateintorobotsfinding their way throughthe physical world:autonomous three-dimensional navigation of dynamic environments requiresa mature software,hardware, and dataecosystemthat simply does not existat scaletoday.NVIDIAstatespart of theproblem plainly:“Large language models are one-dimensional, able topredict the next token, in modes like letters or words. Image-and video-generationmodels are two-dimensional, able to predict the next pixel. None of these models canunderstand or interpret the3Dworld.”8 Theprimary challengesfacing Physical AIare the sameonesthathavetroubled therobotics industry for generations:technology barriersandeconomic barriers. Parts oftheroboticssupply chainremainin theirindustrial infancy,key hardware technologybreakthroughs remain elusive,andeven recentadvancesarenot ready forscalablemanufacturing.Batteries, motors, sensors, and actuators evolvefar more slowly than Center for Security and Emerging Technology |1 algorithmsand software, andscalable manufacturingrequires large amounts ofpatient capital.In addition,much ofthesupply chainfor robotics componentsiscommoditized,and the relatively slim margins dissuade innovative startupsfromcompeting with established incumbents.Adding to these challenges, eachroboticscompany is pursuing its own unique approach, meaning the supply chain ofcomponents and parts remains largely non-standardized, hampering scalabilityandaddingcost.The gap between impressive demonstrations in controlled environmentsand the promise of millions ofaffordablerobots acting independently as they navigatethe world is enormous. The focus of this paper isoncharacterizingtheconvergence ofPhysical AIandrobotics, its underlying supply chain, and identifyingcompetitive advantagesas well asconstraints.This paper provides background on the technologyand describes theecosystem and supply chainof hardware and software supplierssupporting thetechnology. It thencharacterizes competitivenessworldwideusing bibliometrics,patents,investment data,andindustry reportsto determine firm leadership,constraints, and breakthroughs acrossthe technology ecosystem from AI foundationmodels and software to hardware component and robot manufacturers as well as endusers.It concludes with a summary of drivers and positive trends,as well asconstraints and limiting trendswith an eye towards opportunities policymakersinterested in promotingthe tech industry’s next breakthroughmomentcanconsider. This paperbuildsonprevious CSET researchlookingat the robotics patent landscapeto characterize competitivenessusing CSET’sMap of Scienceand separate researchthat proposed a methodology for identifying and characterizing an emergingtechnology.9Itconcludes by introducing a template that could be used by policymakersinterested in global competitiveness assessment ofotheremerging technologies. Table of Contents Executive Summary................................................................................................................................1Introduction...............................................................................................................................................4Scoping and Defining the AI-R