内容来源：https://aibusiness.com/automation/physical-ai-moves-beyond-traditional-robotics

内容总结：

生成式AI的落地应用正从软件走向物理世界。企业开始将AI赋能的机器人系统引入数据中心、制造工厂和仓库等场景，推动“物理AI”从传统机器人向更灵活、更具适应性的方向演进。

从执行指令到理解目标

传统工业机器人通常在受控环境中执行预设指令，而物理AI在此基础上增加了感知与决策能力，使机器人能够应对未明确编程的情况。Gartner分析师比尔·雷将这一转变描述为“从告诉机器人如何完成任务，到给它一个目标、让它自行决定如何实现”。例如，不再指令机器人“拿起箱子放到那里”，而是告诉它“箱子需要装进货车”。不过，雷也提醒，这一方式仍处于早期阶段。

世界模型可帮助机器人在行动前预测物理环境对不同动作的反应，从而支撑这种适应性。但雷表示，机器人自主确定实现目标的方式并适应变化条件的能力“虽已在演示和部分部署中显现，但尚不足以可靠应对现实世界中的全部情况”。

AI视觉与仿真训练

AI驱动的视觉技术正帮助机器人变得更加灵活。ABB机器人业务线行业董事总经理克雷格·麦克唐奈表示，AI正推动工业自动化“从静态、预编程的工位走向自主、可适应、多功能的机器人”。机器人可利用视觉识别产品、检测缺陷、定位零件，并据此行动。

ABB的RobotStudio HyperReality系统利用数字孪生技术，让制造商在模拟环境中训练机器人，涵盖不同光照、材料、颜色和角度等变化，从而在物理部署前验证AI模型。雷举例说，一项传统自动化难以完成的制造任务——机器人在测试站之间移动屏幕，下一步去向取决于上一步测试结果——通过仿真开发耗时四个月，但部署仅用36小时，且对现有流程干扰极小。

数据中心、工厂与仓库的实践

Meta正在多个数据中心测试机器人执行技术人员日常工作，如更换网线和重新安装硬件组件。在爱荷华州阿尔图纳园区，双臂机器人已练习网线更换；在俄亥俄州新奧尔巴尼的Prometheus设施中，Meta正在测试机器人重新安装硬件组件，并探索机械臂能否对服务器进行断电重启。

汽车制造商日产在田纳西州士麦那工厂部署AI驱动的自主移动机器人运输零部件，DHL供应链等物流运营商则使用机器人引导仓库拣货员。

仓库场景中，Robust.AI联合创始人兼首席创新官安东尼·朱尔斯指出，“仓库和物料搬运中90%的活动由带轮设备完成”，因此人形机器人并非必要。该公司的机器人与DHL供应链的仓库员工协同工作，在货架间移动并引导工人拣货。

部署挑战：集成、安全与信任

企业机器人部署远不止让机器完成任务。朱尔斯指出，企业机器人需要两类集成：一是连接ERP、仓库管理和制造执行等系统，二是“变革管理”——缺乏支持的变革会导致采用缓慢甚至失败。雷建议企业应让机器人适应现有流程，而非围绕机器人重新设计流程，因为未来五年机器人本身将发生重大变化。

安全是另一关键挑战。雷强调，随着机器人与人类共同工作，防护设施（如笼子）的成本可能超过机器人本身。他预计更多机器人将内置安全限制，如降低速度或力度，以确保即使出错也不会造成严重伤害。

硬件本身也限制机器人的可靠性。雷指出，人类指尖在感知力和压力方面极为精密，复制这一能力非常困难，“即使让机器人指尖维持三个月也是一项重大成就，但这仍不足以满足企业实际使用需求”。

朱尔斯还警告“新奇陷阱”：企业购买少量机器人做实验却无明确投资回报。“很多所谓的部署实际上只是试点。”

最后，物理AI还带来信任要求。朱尔斯表示，物理系统需要的信任与软件不同，因为“它们就在人旁边工作，任何这样的东西，你都需要去触摸、感受并真正理解”。

中文翻译：

由谷歌云赞助
选择你的首批生成式AI用例
要开始使用生成式AI，首先应关注那些能够改善人类信息体验的领域。
物理AI正在超越传统机器人技术，进入真正的企业任务领域，但企业在整合和扩展机器人系统方面仍面临巨大挑战。
随着企业在数据中心、制造工厂和仓库中测试并使用AI赋能的机器人系统，物理AI正在传统机器人技术之外找到新的应用场景。
这些系统将机器人技术与AI能力相结合，例如计算机视觉、决策制定和世界模型——即旨在表征物理环境并预测其在不同动作下可能如何变化的AI模型。
例如，Meta正在其多个数据中心测试机器人，用于传统上由技术人员执行的 tasks，如更换网线和重新安装硬件组件。在这家社交媒体巨头位于爱荷华州阿尔图纳的园区，双臂机器人一直在练习更换网线。在其位于俄亥俄州新奥尔巴尼较新的Prometheus设施中，Meta正在测试用于重新安装硬件组件的机器人，同时还在探索机械臂是否能够对服务器进行断电重启。
类似的应用正在其他行业涌现。汽车制造商日产正在其位于田纳西州士麦那的工厂部署AI驱动的自主移动机器人来运输零部件，而DHL供应链等物流运营商正在使用机器人引导工人完成仓库拣货。
虽然物理AI的应用范围十分广泛，但它们有一个共同的务实焦点：用机器人来完成重复性、体力要求高或难以通过人力扩展的工作。
机器人在企业中已经工作了几十年。传统工业机器人通常在受控环境中执行预定义指令。物理AI增加了感知和决策能力，使机器人能够应对它们未被明确编程处理的情况。
Gartner分析师Bill Ray将这一转变描述为从告诉机器人如何执行任务，转变为给它一个目标，让它自己决定如何实现。一个例子是告诉机器人箱子需要被放进货车里，而不是指示它拿起箱子并放到那里。不过，Ray警告说，这种方法仍处于早期阶段。
世界模型可以通过帮助机器人在行动之前预测物理环境在不同动作下可能如何变化，来支持这种适应性。
他说：“机器人确定如何实现目标并适应变化条件的能力正在演示和一些部署中显现，但它还不够可靠，尚不足以应对现实世界中的全部条件。”
AI驱动的视觉是帮助机器人变得更具适应性的一项技术。ABB机器人业务线工业部门董事总经理Craig McDonnell表示，AI正在推动工业自动化“从静态、预编程的工位转向自主、自适应和多用途的机器人”。机器人可以利用视觉来识别产品、检测缺陷和定位零件，然后根据所看到的内容采取行动。Meta正在其Prometheus数据中心测试ABB机器人的机器人。
但这并不意味着每个AI赋能的机器人都是自主的。人工监督仍然普遍存在，尤其是在昂贵设备周围，但更强的感知能力和适应性可以扩大机器人能够处理的任务范围。
物理AI也不一定意味着人形机器人。Ray使用“多功能机器人”一词来描述更通用的机器，其功能可以通过软件改变。他说，在企业环境中，机器人在相对结构化的空间里搬运物料时，轮子往往比腿更有意义。他指出，腿式机器人必须不断保持平衡，这使其能耗很高。
同样，人形机器人对企业任务来说并不总是必要的，Robust.AI联合创始人兼首席创新官Anthony Jules说道。Robust.AI是一家为仓库开发AI驱动协作移动机器人的机器人公司。他说：“仓库和物料搬运中90%的活动都是由带轮子的设备完成的。”
物理AI正在整个企业中找到实际应用，尽管大多数部署仍集中在特定任务上。早期案例涵盖数据中心、工厂和仓库，企业在这些场所使用机器人从事重复性、高要求或难以通过人力扩展的体力工作。
Meta的实验表明，企业正在将机器人技术应用于AI基础设施运营，而非传统的制造或仓库工作。较新的试点涉及机器人直接与数据中心设备交互，包括网线和服务器硬件。
Ray表示，数据中心对物理AI来说可能是有吸引力的环境，因为许多维护任务具有重复性，而且设施可以提前绘制地图。与此同时，服务器机房包含密集的线缆、不同的设备配置以及在同 一空间中工作的技术人员，这给机器人带来了比传统自动化通常处理的更多的可变性。
工厂长期以来一直在使用机器人，但通常只用于一项固定工作。ABB机器人的McDonnell表示，物理AI超越了“每个机器人只做一项工作的传统局限”。他说，物理AI的一些最大机会在于那些由于高可变性而传统上难以自动化的应用，包括混合产品、处于不同位置和方向的组件，以及不同的照明或环境条件。
ABB正在利用AI驱动的仿真来应对其中一些挑战。其RobotStudio HyperReality系统让制造商能够在模拟工厂环境中训练机器人，并引入照明、材料、颜色和角度的变化。该方法依赖数字孪生——产品、机器人和操作环境的虚拟副本，使制造商能够在实体部署之前，在多种条件下训练和验证AI模型。
仿真还可以帮助机器人在物理世界中行动之前做出决策。Gartner的Ray描述了一个制造案例，其中机器人要在测试站之间移动屏幕，而每一个下一站都取决于前一个测试结果。这项任务对传统自动化来说过于复杂。他说，使用仿真进行开发花了四个月，但部署只用了36小时，并且对现有流程的干扰极小。
仓库长期以来一直使用机器人来搬运库存，但较新的系统正在加入AI，以帮助机器人处理更多可变任务，并与员工协同工作。亚马逊已在其履约网络中构建了大规模机器人基础设施，使用移动机器人运输库存，并利用AI协调其机器人队伍。
Ray说，仓库已经在使用先进的专用机器人，因此通常几乎没有理由用人形机器人取代它们。相反，价值在于找到灵活机器人能够适应现有流程的任务。
其他部署则侧重于将机器人整合到现有仓库工作流程中。Robust.AI与DHL供应链的合作提供了一个更具渐进性的例子。Robust.AI的Jules说：“我们的机器人与仓库员工并肩工作，通过在不同位置之间移动并引导工人前往何处来帮助他们拣货。”
这一部署说明了如何在不要求企业替换更广泛工作流程的情况下，将物理AI添加到现有运营中。但要让这种方法在生产中奏效，企业必须处理好机器人周围的系统、流程和人员。
对企业机器人而言，让机器执行任务只是部署的一部分。机器人还必须与现有IT系统连接，并融入它所协同工作的人类工作流程。
Jules说，企业机器人需要两种整合：将机器人连接到ERP、仓库管理和制造执行等系统，以及第二个往往更困难的部分：“变革管理部分。”他说，支持不力会导致采用缓慢甚至无法采用。
这使得机器人对现有运营的适配性与它的技术能力同样重要。Ray说，企业应让机器人适应现有流程，而不是围绕机器人重新设计流程，因为机器人在未来五年内将发生重大变化。
一旦机器人适配了工作流程，它仍然必须足够可靠和安全，才能用于生产。McDonnell说：“在要求严苛的工业和商业环境中，性能哪怕下降一点点都是不可接受的。”
Ray强调，随着机器人在人类周围工作，安全变得更加重要，而防护基础设施（如笼子）的成本可能比机器人本身还高。他预计会有更多机器人被设计为内置安全限制，例如较低速度或减小力度，这样即使出现问题，也不会严重伤害到人。
安全只是技术挑战的一部分。硬件本身也可能限制机器人能够可靠完成的任务。
Ray说：“人类指尖在感知力和压力方面极其精密，复制这一点非常困难。哪怕让一个机器人指尖持续使用三个月，也可能是一项重大成就，但这对于实际企业使用来说仍然不够长。”然而，技术可靠性并不能保证企业部署在经济上合理。Jules将此描述为“新奇陷阱”：公司买几个机器人做实验，却没有明确的投资回报率。他说：“很多人们称之为部署的东西，实际上只是试点。”
除了技术和财务考量之外，物理AI还带来了另一项要求。
那就是信任。
Jules说，物理系统需要的信任与软件不同，因为“它们就处在人们身边并与人们一起工作，任何类似这样的东西，你都需要去触摸、去感受、去真正理解。”

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
Physical AI is moving beyond traditional robotics into real enterprise tasks, but enterprises still face big challenges integrating and scaling robotic systems.
Physical AI is finding new applications beyond conventional robotics as enterprises test and use AI-enabled robotic systems in data centers, manufacturing plants and warehouses.
These systems combine robotics with AI capabilities, such as computer vision, decision-making and world models -- AI models designed to represent physical environments and predict how they might change in response to different actions.
Meta, for example, is testing robots at several of its data centers for tasks traditionally performed by technicians, such as swapping network cables and reseating hardware components. At the social media giant's Altoona, Iowa, campus, dual-armed robots have been practicing network-cable swaps. At its newer Prometheus facility in New Albany, Ohio, Meta is testing robots to reseat hardware components, while also exploring whether a robotic arm can power-cycle servers.
Similar applications are emerging in other industries. Automaker Nissan is deploying AI-powered autonomous mobile robots to transport components in its Smyrna, Tenn., factory and logistics operators, such as DHL Supply Chain, are using robots to guide workers through warehouse picking.
While applications of physical AI vary widely, they share a practical focus: using robots for work that's repetitive, physically demanding or hard to scale with human labor.
Robots have worked at enterprises for decades. Traditional industrial robots typically execute predefined instructions in controlled environments. Physical AI adds perception and decision-making capabilities, enabling robots to respond to conditions they weren't explicitly programmed to handle.
Bill Ray, an analyst at Gartner, described the shift as moving from telling a robot how to perform a task to giving it a goal and letting it determine how to achieve it. An example of this is telling a robot that boxes need to be in a van rather than instructing it to pick up a box and put it there. However, Ray cautioned that this approach is still in the early stages.
World models can support this kind of adaptability by helping robots predict how the physical environment might change in response to different actions before they act.
"That ability of robots to determine how to achieve a goal and adapt to changing conditions is emerging in demonstrations and some deployments, but it isn't yet reliable enough to handle the full range of real-world conditions," he said.
AI-powered vision is one technology helping robots become more adaptable. Craig McDonnell, managing director, business line industries, at ABB Robotics, said AI is shifting industrial automation "from static, preprogrammed stations toward autonomous, adaptable and versatile robots." Robots can use vision to recognize products, detect defects and locate parts, then act on what they see. Meta is testing ABB Robotics' robots at Meta's Prometheus data center.
But that doesn't mean every AI-enabled robot is autonomous. Human oversight remains common, especially around expensive equipment, but greater perception and adaptability can expand the range of tasks robots can handle.
Physical AI also doesn't necessarily mean humanoid robots. Ray used the term polyfunctional robots to describe more general-purpose machines whose functions can change using software. He said wheels often make more sense than legs in enterprise environments where robots move materials across relatively structured spaces. Legged robots must constantly balance, he noted, which makes them power-intensive.
Similarly, humanoids aren't always necessary for enterprise tasks, said Anthony Jules, co-founder and chief innovation officer at Robust.AI, a robotics company developing AI-powered collaborative mobile robots for warehouses. "Ninety percent of activity in warehouses and material handling is done by things with wheels," he said.
Physical AI is finding practical applications across the enterprise, although most deployments still focus on specific tasks. Early examples span data centers, factories and warehouses, where companies use robots for physical work that's repetitive, demanding or difficult to scale with human labor.
Meta's experiments show how companies are applying robotics to AI infrastructure operations rather than conventional manufacturing or warehouse work. The newer pilots involve robots interacting directly with data center equipment, including network cables and server hardware.
Ray said data centers can be attractive environments for physical AI because many maintenance tasks are repetitive and facilities can be mapped in advance. At the same time, server halls contain dense cabling, varied equipment configurations and technicians working in the same space, creating more variability for robots than conventional automation typically handles.
Factories have long used robots but usually for one fixed job. Physical AI moves beyond "the traditional limitation of 'one job per robot," ABB Robotics' McDonnell said. Some of the greatest opportunities for physical AI lie in applications that have traditionally been hard to automate because of high variability, he said, including mixed products, components in different positions and orientations, and varied lighting or environmental conditions.
ABB is addressing some of those challenges using AI-powered simulation. Its RobotStudio HyperReality system lets manufacturers train robots in simulated factory environments with variations in lighting, materials, colors and angles. The approach relies on digital twins, virtual replicas of products, robots and operating environments, so manufacturers can train and validate AI models in many conditions before physical deployment.
Simulation can also help robots make decisions before they act in the physical world. Gartner's Ray described a manufacturing example in which robots move screens between test stations, with each next stop depending on the previous test result. The task was too complex for conventional automation. Development using simulation took four months, but deployment took 36 hours with minimal disruption to the existing process, he said.
Warehouses have long used robots to move inventory, but newer systems are adding AI to help robots handle more variable tasks and work alongside employees. Amazon has built large-scale robotic infrastructure across its fulfillment network, using mobile robots to transport inventory and AI to coordinate its fleet.
Warehouses already use advanced specialized robots, so there's often little reason to replace them with humanoid ones, Ray said. Rather, the value comes from finding tasks for which flexible robots fit existing processes.
Other deployments focus on integrating robots into existing warehouse workflows. Robust.AI's work with DHL Supply Chain offers a more incremental example. "Our robots work alongside warehouse employees, helping them pick orders by moving between locations and directing workers where to go," Robust.AI's Jules said.
The deployment illustrates how physical AI can be added to an existing operation without requiring companies to replace the broader workflow. But making that approach work in production requires companies to address the systems, processes and people around the robot.
For enterprise robotics, getting a machine to perform a task is only part of the deployment. The robot also must connect with existing IT systems and fit into the human workflows it's working alongside.
Enterprise robotics requires two kinds of integration, Jules said: linking the robot to systems such as ERP, warehouse management and manufacturing execution, and a second, often harder piece: "the change management piece." Poorly supported change leads to slow or no adoption, he said.
That makes the robot's fit within the existing operation just as important as its technical capabilities. Ray said enterprises should adapt robots to existing processes rather than redesigning processes around them, since robots will change substantially over the next five years.
Once a robot fits the workflow, it still must perform reliably and safely enough for production. "Even a fractional drop in performance is unacceptable in demanding industrial and commercial environments," McDonnell said.
Ray emphasized that safety grows more important as robots work around people and that protective infrastructure, such as cages, can cost more than the robot themselves. He expects more robots to be designed with built-in safety limits, such as lower speeds or reduced force, so they can't seriously injure someone even if something goes wrong.
Safety is only part of the technical challenge. The hardware itself can limit what robots can reliably do.
"Human fingertips are incredibly sophisticated at sensing force and pressure, and replicating that is really difficult," Ray, said. "Even getting a robotic fingertip to last three months can be a significant achievement, but that's still not long enough for practical enterprise use." Technical reliability, however, doesn't guarantee that an enterprise deployment will make financial sense. Jules described this as a "novelty trap": companies buying a few robots to experiment without a clear ROI. "A lot of things that people call deployments are actually just pilots," he said.
Beyond the technical and financial considerations, physical AI introduces another requirement.
That's trust.
Physical systems demand a different kind of it than software, Jules said, because "they're situated and working next to people, and anything that's like that, you need to touch and feel and really understand."