Robots made frequent stunning appearances at this year’s CCTV Spring Festival Gala, presenting wonderful performances in succession. Beneath the stage lights, the audience could clearly feel that artificial intelligence has become tangible and perceivable in the physical world. AI robots are undergoing a critical transition from laboratory research to real-world application, evolving from conversational tools into practical performers. This transformation is driven by a cutting-edge advancement in artificial intelligence — physical AI.
Physical AI was also a trending topic at the 2026 Consumer Electronics Show (CES) held in Las Vegas, the United States. As a bellwether of the global consumer electronics industry, CES serves as a vital window to observe technological trends. Compared with AI agents that dominated discussions in 2025, the spotlight has shifted to physical AI this year, reflecting changes in AI development paths and industrial priorities.
What exactly is physical AI? What changes will it bring? What lies ahead for its development, and what challenges remain? Let’s take an in-depth look.
Starting with a Famous Paradox in Technology
Artificial intelligence has penetrated all industries and daily life. From early search engines and voice assistants to modern large language models such as ChatGPT and DeepSeek, people have grown accustomed to AI’s capacity for language comprehension and content creation. These models can compose articles, draw pictures and even write codes. Nevertheless, AI often struggles with routine housework, frequently knocking over cups or bumping into furniture.
This seeming contradiction is known as Moravec's Paradox, a renowned anomaly in the AI field.
Back in the 1980s, AI pioneer Hans Moravec discovered an intriguing phenomenon: tasks challenging for humans like playing Go and solving calculus problems come easily to robots. By contrast, instinctive human actions such as walking and grasping objects pose immense difficulties for machines.
Traditional AI is essentially a disembodied mind. It can analyze lexical connections from massive datasets, describe gravity elegantly and formulate physical laws mathematically. Yet it cannot perceive that a glass will shatter when dropped, nor recognize slip hazards on wet floors. Confined to the digital realm of binary codes, it lacks the ability to sense and interact with the physical world governed by friction, inertia and uncertainties.
Decades of technological evolution have validated this theory. In 1997, Deep Blue defeated a world chess champion. In 2016, AlphaGo outmastered Go player Lee Sedol. By 2023, large models were capable of academic writing, programming and hyper-realistic image generation. On the flip side, numerous humanoid robots still stumbled and malfunctioned frequently at the 2025 World Humanoid Robot Games. Even unscrewing a bottle cap proved a daunting task.
Simply put, conventional AI excels at verbal expression but falls short of practical execution. Against this backdrop, physical AI has emerged as a breakthrough frontier.
In simple terms, physical AI integrates artificial intelligence with the tangible physical world. Moving beyond theoretical operations on screens, it empowers robots to move and interact freely in real-life scenarios. Agile quadruped robots, high-precision warehouse sorting robotic arms and household vacuum cleaners all leverage physical AI technology.
Tang Jie, founder of Zhipu AI and professor at Tsinghua University, states that the era of conversational AI is drawing to a close, and the industry is stepping into a new phase of action-oriented AI.
Expanding Boundaries: From Virtual Information to Physical Execution
Physical AI refers to intelligent systems that perceive surroundings, interpret physical principles, plan motions and accomplish complex real-world tasks. Instead of delivering theoretical answers, it focuses on tangible outcomes, such as opening doors, transporting cargo and traversing rugged terrain steadily.
AI development can be divided into four stages. The initial perception AI enables visual and auditory detection. It is followed by generative AI that produces texts and images. Agent AI can invoke and combine diverse software tools. The newly emerging physical AI understands real environments and carries out practical operations.
Unlike traditional AI, physical AI features a tangible physical form. It collects environmental data via cameras, lidar and tactile sensors, and executes physical movements through motors and joints. The whole process forms a dynamic loop: perception, cognition, action and re-perception.
It also boasts powerful spatial awareness. Beyond basic object recognition, it analyzes three-dimensional positional relationships. It can judge whether a vase will fall off a table edge, and distinguish material properties to determine proper force for lifting boxes.
Physical AI represents the combined application of embodied intelligence and spatial intelligence. Embodied intelligence provides theoretical frameworks, while spatial intelligence underpins practical task execution. By merging these capabilities into operable systems, physical AI creates functional intelligent entities.
AI is steadily breaking free from virtual limitations and stepping into physical reality, marking an evident technological inflection point.
Unlocking Unlimited Real-world Application Potential
Conventional robots run rigidly on preset programs, repeating fixed routes, grasping designated items and performing set movements. They fail to adapt once surroundings change. Robots equipped with physical AI gain unprecedented flexibility and adaptability, accelerating industrial adoption.
More Efficient Industrial Manufacturing: With relatively stable operating conditions, industrial robots have been widely deployed on production lines. Tesla’s welding robots achieve precision within 0.1 millimeters with physical AI assistance. In a factory in Mianyang, Sichuan, autonomous robots shuttle smoothly to transport containers. They dodge obstacles, predict workers’ paths and yield voluntarily, behaving like seasoned staff members.
More Accurate Medical Assistance: Next-generation surgical robots evolve from auxiliary tools into reliable medical partners. They analyze tissue elasticity and blood flow in real time, adjusting suture strength dynamically. Clinical data shows that da Vinci surgical robots powered by physical AI cut intraoperative blood loss by 40%. Ultrasound puncture robots also see drastically reduced error rates after virtual organ simulation training.
More Reliable Autonomous Driving: Smart vehicles achieve substantial upgrades. Early driving assistance systems rely on road signs, lane lines and surrounding vehicles for decision-making. Physical AI enables cars to comprehend physical dynamics, identifying icy road surfaces, forecasting electric vehicle maneuvers and judging pedestrian crossing intentions. One automaker reports that its upgraded driving system boosts average manual takeover distance by 13 times on complex mixed-traffic roads.
Intimate Household Services: Household robots gain enhanced practicality. Vacuum cleaners automatically detour around slippers and cables, and switch cleaning modes according to floor materials and dirt levels. Advanced robots with mechanical arms can tidy toys, deliver medicine via wheelchairs and assist the elderly in standing up. These functions stem from 3D environmental perception, material and weight recognition, and human intention deduction. Household robots will grow increasingly versatile and become practical daily helpers.
Bottlenecks Hindering Widespread Popularization
Despite promising prospects, physical AI still faces multiple hurdles before full-scale implementation across industries and households.
Data acquisition poses the primary challenge. Autonomous driving systems require millions of kilometers of road test data to cover emergency scenarios, incurring huge training costs. Robotic arms undergo countless trial runs to handle delicate components gently. Moreover, physical laws vary drastically across sectors, making universal training datasets inapplicable for surgical and industrial robots. Researchers turn to synthetic data generated by virtual simulation, aiming to transfer virtual training results to real-world scenarios.
This gives rise to the second obstacle: the gap between simulated and actual environments. Robots perform flawlessly in virtual simulations yet malfunction in reality. For instance, an agent can grasp apples effortlessly in digital scenes but may drop them due to uneven surfaces or water droplets. Bridging this simulation-to-reality gap remains a core research priority.
Furthermore, physical AI constitutes an intricate systematic project. Its performance cannot be improved merely by expanding computing power and data volume like large language models. It integrates artificial intelligence, mechanical engineering, signal processing and other disciplines, with flaws in any link undermining overall performance.
Safety risks also cannot be overlooked. Physical AI controls solid mechanical equipment with mass and force. Erroneous judgments may lead to device damage, property loss and even personal injury, demanding higher reliability standards.
A consensus has emerged amid the booming physical AI trend: AI’s future lies far beyond text and content generation. It will perceive, understand and reshape the physical world. Serving as the key to resolving Moravec's Paradox, physical AI equips thinking-oriented AI with dexterous and dependable operational capabilities, ushering artificial intelligence into a brand-new developmental era.
Source: PLA Daily, Military News Network of China
Author: Guo Ning
从“会聊天”到“会做事”，一起来关注物理AI
在今年中央广播电视总台春节联欢晚会上，机器人高密度登场，轮番献上精彩的节目。聚光灯下，台下观众直观地感受到，如今AI在真实物理世界中可触可感。从“会聊天”到“会做事”，AI机器人正在经历从“实验室”到“应用端”的关键转折。这背后，离不开人工智能技术发展的最新技术——物理AI。
在前不久举办的2026年美国拉斯维加斯消费电子展（CES）上，“物理AI”也被频频提及。作为国际消费电子领域的风向标，CES历来是观察前沿技术走向的重要窗口。与2025年热议的“AI智能体”相比，今年展会热议“物理AI”这一新词，折射出AI技术发展路径和产业关注重点的变化。
那么，什么是物理AI？它将带来哪些改变？未来发展前景如何？又面临哪些挑战？请看本期关注。
从一个著名的“倒挂”现象说起
当今社会，人工智能浪潮席卷千行百业，越来越融入人们的日常生活。从最早的搜索引擎和语音助手，到近几年的ChatGPT、DeepSeek等大语言模型，人们已经习惯于让AI“理解语言”“生成内容”。这些大模型应用可以写文章、画图，甚至编写程序。然而，当我们想让AI去做家务时，它大概率会手忙脚乱，甚至打翻杯子、撞上桌角。
这听起来很矛盾，却是人工智能领域一个著名的“倒挂”现象——“莫拉维克悖论”。
早在上世纪80年代，人工智能先驱汉斯·莫拉维克就发现：对人类来说，下围棋、解微积分很难，但对机器人来说可能很容易；而走路、抓东西这些人类不假思索就能完成的动作，对机器人而言却异常艰难。
传统AI，本质上是个“脱体的大脑”。它或许能从海量数据中学习词语间的关联，用优美的语言描述“重力”，也能用数学符号写出物理定律公式，却并不能发觉玻璃杯掉在地上会碎，也不会在意湿滑的地面容易让人摔倒。它在由0和1构成的数字世界里，无法感知也无法作用于我们这个充满摩擦力、惯性和不确定性的物理世界。
过去几十年，人工智能领域的发展也印证了这一点。1997年，“深蓝”大模型击败国际象棋冠军；2016年，AlphaGo大模型战胜围棋高手李世石；到了2023年，大模型能写论文、编程序、生成逼真图像……与之相对，在2025年举办的世界人形机器人运动会上，不少机器人仍频频摔倒、“洋相”频出。拧开瓶盖这一简单动作，对它们来说都是一场高难度挑战。
换句话说，过去的AI很会“说话”，却不擅长“做事”。在这一背景下，人工智能的又一技术前沿——物理AI技术迅速发展起来。
所谓“物理AI”，通俗地说，就是将AI与物理世界深度融合的人工智能技术。物理AI不再满足于只在屏幕里“纸上谈兵”，而要让机器人学会在真实世界里自如地行动和互动。从灵活奔跑的机器狗，到仓库里精准分拣包裹的机械臂，再到家用扫地机器人……它们的背后，都有物理AI技术应用的影子。
智谱AI公司创始人、清华大学教授唐杰表示，对话AI的范式基本接近尾声，将进入“做事AI的阶段”。
能力边界从虚拟信息层跨越到真实行动层
换言之，物理AI是一种能够在现实世界中感知环境、理解物理规律、规划动作并执行复杂任务的智能系统。它的目标不是给出一个理论答案，而是完成一件具体的事——比如打开一扇门、搬运一个箱子，或是在崎岖的山路上稳健行走。
有人将AI发展分为4个阶段：初期的“感知AI”，能看能听；随后的“生成AI”，能够输出文字、图像等内容；“代理AI”，能够调用、组合不同的软件工具；现在的“物理AI”，能够理解现实世界并执行具体的操作任务。
与传统AI相比，物理AI最核心的特点在于拥有“身体”。它通过摄像头、激光雷达、触觉传感器等“感官”从真实世界获取信息，并通过电机、关节等“肢体”输出一系列物理动作。整个动作过程是一个“感知—思考—行动—再感知”的动态闭环。
与此同时，物理AI还具备强大的空间智能感知能力。物理AI不仅要识别物体，还要理解三维空间中物体的相对关系：这个花瓶放在桌子边缘，推一下会不会掉？那个箱子是纸做的还是铁做的，该用多大力气搬？
从这一角度来说，物理AI可以看作是具身智能与空间智能融合的应用范式。具身智能提供理论框架，空间智能提供实际执行任务的关键能力，物理AI将这些能力整合到现实可操作的系统中，最终落地为一个个能干活的智能实体。
从虚拟信息层跨越到真实行动层，AI的能力边界不断拓展。一个日见清晰的技术拐点，正呈现在我们面前。
推开一扇扇通往真实世界的“门”
传统机器人往往是“死脑筋”，大多按照预设程序重复单一动作：走规划好的线路、抓取预定位置的物体，以及跳固定动作的舞蹈等。一旦环境变化，它们就会束手无策。而应用物理AI系统的机器人，具有前所未有的灵活性和适应性，落地应用速度明显加快。
——工业制造更加高效。在制造领域，由于应用场景相对稳定，目前车间生产线上的各类工业机器人已经开始大显身手。例如，特斯拉公司推出的焊接机器人借助物理AI辅助，焊接精度突破0.1毫米。位于四川绵阳的某企业车间，多台自主机器人灵活穿行、搬运周转箱，不仅能避开障碍，还会预判工人走动路线，主动让路，俨然一副“老员工”的样子。
——辅助医疗更加精准。在医疗领域，新一代手术机器人将不只是医生可有可无的帮手。它能实时分析人体组织的弹性、血流状况，并自动调整伤口缝合的力度。临床数据显示，搭载物理AI系统的达芬奇手术机器人，能让术中患者出血量减少40%；而用于超声穿刺的机器人，在经过虚拟器官模型训练后，操作失误率更是大幅下降。这些机器人正从饱览医书病历的分析工具，渐渐升级为具备精湛实操技术的“手术搭档”。
——自动驾驶更加可靠。在交通领域，自动驾驶汽车也将迎来重要飞跃。过去的智驾系统主要靠识别路标、车道线和四周车辆辅助司机决策。而物理AI则让汽车开始“理解”物理世界的动态规律，如判断路面是否结冰，预测汽车旁边电动车骑手的下一步动作，甚至与横穿马路的行人进行“眼神交流”式的意图沟通。例如，某汽车公司宣称，其融合物理AI的新一代智驾系统，实现了在复杂小路和人车混行的场景下，平均接管里程提升了13倍。
——家政服务更加贴心。在家政领域，变化同样显著。物理AI让扫地机器人能辨认出地上的拖鞋、电线，主动绕行；面对瓷砖、木地板或地毯等不同材料地面的不同脏污程度，机器人能自动切换清扫模式。更令人惊喜的是，一些同时搭载机械臂的家政机器人，已经能够整理散落玩具、推轮椅送药，甚至协助老人起身。这些能力背后，是物理AI对家庭环境的三维理解、对物体材质与重量的感知，以及对人类意图的推理。可以预见，未来的家政机器人功能将会越来越多，成为真正懂生活、会干活的“家庭帮手”。
走向全面深入应用仍面临诸多挑战
尽管物理AI应用前景令人激动，但其想要真正走进千行百业、走进千家万户，仍面临着诸多挑战。
首先是数据难题。训练一个能在马路上安全驾驶的AI汽车，物理AI系统要积累数百万公里的真实路测数据，才有可能涵盖各种突发情况，而这往往意味着高昂的训练成本；工厂里的机械臂也需要成千上万次试错，才能学会轻拿轻放精密零件。更麻烦的是，不同行业场景需要考虑的物理规律天差地别，例如，手术机器人和工业搬运机器人根本不能采用同一套数据训练。为了应对这一问题，专家们寄希望于合成数据（在虚拟世界中通过仿真技术生成的数据），或者在高度逼真的虚拟世界中训练AI，再迁移到现实世界中。
但这又将带来第2个风险挑战：“仿真世界”和“真实世界”的鸿沟。很多机器人在虚拟环境中被训练得像个“学霸”，可一放到现实里就“挂科”。比如，有的智能体在模拟环境中能轻而易举地抓一个苹果，但现实环境中，由于苹果表面可能有水珠、或者苹果形状不规则等，稍有动作偏差，苹果就会从机械手中滑落。如何缩小从仿真到现实的差距，也是AI工程师们未来重点攻关的课题。
此外，物理AI是一个高度复杂的系统工程，它不像大语言模型那样，依靠堆叠算力和数据就能提升能力。物理AI融合了人工智能、机械工程、信号传感处理等多个领域，任何一个环节的短板都会影响系统整体表现。
最后，安全问题也不容小觑。物理AI系统操控的是有质量、有力量的实体设备。一次错误的决策，就可能导致设备损坏、财产损失，甚至造成人身伤害。因此，未来物理AI要更加可靠。
面对汹涌而至的物理AI浪潮，一个共识正在形成：AI的未来，不再只是生成华丽的词藻和内容，更要真正进入世界、理解世界、改造世界。物理AI，正是破解“莫拉维克悖论”的钥匙，它正为那个曾经只会“思考”的大脑装上一双灵巧、可靠的手，推动人工智能发展步入一个崭新的阶段。
来源：中国军网-解放军报 作者：郭 宁
返回首页