Can robots understand the world?

Which task is harder: teaching a robot to do a backflip, or teaching it to fold a piece of clothing? The answer is the latter. A backflip tests "understanding one's own body," while folding clothes tests "understanding the world." Making robots understand the world is a genuinely difficult problem. In the world of AI, humanoid robots and embodied intelligence are new concepts in the field of artificial intelligence. What is embodied intelligence? How should we understand "embodied"? Given that we already have traditional robots, why do we still need to build humanoid robots? What Is Embodied Intelligence In the field of artificial intelligence, one term that has become very popular is embodied intelligence. What exactly is embodied intelligence? Before discussing embodied intelligence, let us first introduce a corresponding concept: disembodied intelligence. Disembodied intelligence excels at processing abstract information such as text and images and performing logical reasoning; traditional large language models belong to disembodied intelligence. The training material for disembodied intelligence is internet text and images, and it learns the statistical associations between words. It can output the words "unscrew the bottle cap," but it does not know how slippery the cap is, how much torque is required, or how the hand must steady the bottle. To it, words are merely symbols, lacking any grounding in the physical world. Embodied intelligence means empowering AI onto a physical entity, enabling AI to interact with the real physical world. Cognitive science long ago proposed the concept of "embodied cognition." Human intelligence does not exist solely in the brain; our understanding of weight, softness and hardness, and balance all comes from decades of bodily experience. Infants build common-sense understanding of the world by grabbing things, falling down, and touching things hot and cold. Traditional AI in the past skipped this process, whereas what embodied intelligence seeks to do is replicate this path of "learning common sense through practice." Many people conflate embodied intelligence with robots and humanoid robots. A considerable portion of embodied intelligence is indeed presented with robots as the carrier, but the two cannot be equated. Embodied intelligence is not entirely equivalent to humanoid robots; its carriers can also be self-driving cars, inspection drones, and so on. Zhang Xueting, project director of the Embodied Intelligence Experimental Class at Hangzhou Dianzi University and associate professor in the School of Mechanical Engineering, points out that the core criterion for distinguishing whether a robot is embodied intelligence is that embodied intelligence relies on environmental perception for self-evolution and iterative learning. Traditional robots, although equipped with sensors, preset programs, and a certain degree of execution capability, do not necessarily qualify as embodied intelligent robots. For example, some science museums have humanoid robots with highly realistic appearances. When you talk to them, they will perform basic actions such as raising a hand or blinking, but all their answers and actions are pre-programmed. If they encounter a question outside the settings, they cannot answer it. Even if their appearance resembles a real person, they still only belong to humanoid robots in the broad sense. True embodied intelligent robots have models and programs that rely on environmental perception to achieve self-learning and self-training, continuously iterating and becoming increasingly smart. For example, some humanoid robots currently participating in robot games see their capabilities iterate every month, not because programmers continuously rewrite code on a large scale, but because of the autonomous iterative capability brought by embodied intelligence. Why Robots Are Built to Look Human Some people ask: since we already have traditional robots, why build humanoid robots? Factory robotic arms are precise and efficient, AGVs (automated guided vehicles) can transport goods, and robot vacuum cleaners can clean rooms. These mature robots have solved a great deal of physical labor, so why still develop humanoid robots that walk unsteadily and cost a fortune? In the view of scientists, traditional robots transform the world, making the world adapt to machines; the goal of humanoid robots is to transform machines, making machines adapt to the human world that has already been built. All current infrastructure is designed according to the human form—for example, the height of door handles, stair steps, tables and chairs, shelves, wrenches, switches, and so on. For a robotic arm to take up a post, the workstation must be specially modified and protective fences installed; AGVs can only travel on flat ground along preset routes and struggle to move when encountering steps or narrow doorways. The limbs, torso, and hands of humanoid robots are close in size to those of humans, so in theory they can directly use all the tools humans currently have and enter rooms and warehouses that have not been specially modified. Therefore, the humanoid form is not "to look like a human" but the greatest common denominator of humans' existing physical environment. It is not a mere biomimetic gimmick but an engineering choice made after weighing trade-offs. At an "embodied intelligent robot shop" on Lei Street in Hefei, a robot grabs goods and hands them to a customer. We often see videos of humanoid robots jogging, going up and down stairs, and standing up on their own after falling. Some people believe that at this rate, robots will soon be able to handle all household chores. However, the development of humanoid robots is still far from reaching the stage of mastering humans' "general capabilities." The shortcomings are concentrated in areas such as fine and flexible manipulation, physical common sense, long-task planning, and battery life and durability. For example, what the industry calls the "last few centimeters problem"—dexterous hands. Robots can pick up hard boxes, but it is very difficult for them to stably complete tasks such as unscrewing bottle caps, plugging in plugs, folding soft clothing, or picking up eggs. They lack distributed tactile perception similar to human skin and cannot finely control force. Another example is the insufficient physical common sense of humanoid robots. Large language models have read tens of millions of sentences about water and cups, but they do not intuitively understand physical common sense such as "a cup falls over at a touch" or "water flows." They lack the intuitive world model that humans accumulate in childhood, and they easily get stuck when encountering unexpected situations. Humanoid robots have already achieved "being able to act," but there is still a long way to go before they become general intelligent agents capable of flexible response and autonomously completing a whole set of complex, long tasks. Several hurdles stand between demonstration and practical use. Over the past decade, the core of the AI wave has been processing digital information such as text, images, and audio, while embodied intelligence opens the next leg of the journey, taking AI off the screen and into the physical world, where it learns the real world's causality and common sense through action. So what hurdles must embodied intelligent robots clear before they move from demonstration to practical use? The first major hurdle is the insufficient reliability of robots' "brains." The "brains" of embodied intelligent robots perform passably in preset environments, but once they enter the unstructured real world they easily "drop the ball." Robots struggle to transfer skills learned in one scenario to a new environment, and this is the core root of the "brain" not being smart enough. Wang Xingxing, founder of Unitree Robotics, offers a rather distant outlook for embodied intelligence's "ChatGPT moment"—completing 80% of tasks in 80% of unfamiliar scenarios through natural instructions—at the fastest in 2–3 years, at the slowest in 5–10 years. Unlike language large models, which have massive amounts of internet text available, robots' physical interaction data cannot be obtained "just by reaching out," and the data hurdle is the second major difficulty. Simulated data is low-cost but cannot fully replicate real physical laws (such as materials and friction), while real-machine data is authentic but extremely costly to collect and inefficient. In addition, manufacturers' hardware interfaces and data formats each form their own systems, making data hard to reuse and share and further aggravating the scarcity of high-quality "physical world data." Body hardware "holding things back" is the third major hurdle. Body hardware is the physical foundation of embodied intelligence, and its performance under real conditions and manufacturing capacity remain hard constraints. Under the premise of limited device weight, robots must meet customers' baseline performance requirements for industry applications such as load capacity, high- and low-temperature adaptability, and continuous working hours. The mass-production hurdle is the fourth major difficulty on embodied intelligence's path to practical use. At present, humanoid robot products from many companies at home and abroad have already achieved mass production, except that most of these robots have flowed to scenarios such as education and research and exhibitions, and have not yet truly "taken up posts and worked" in practical scenarios such as factories, logistics, handling, and housework. Industry insiders believe that for embodied intelligence to achieve breakthrough development, interdisciplinary crossover is needed, and comprehensive talents must work as a team to tackle key problems under top-level planning. This "turn" of artificial intelligence still has much room for imagination and development. (Image source in this article: Xinhua News Agency. Cover image: At the Beijing Humanoid Robot Innovation Center, a humanoid robot makes a "finger heart" gesture.)