Storyline: Humanoid robots in China honing home skills for mass domestic adoption [Voice_over] This is a humanoid robot training center. The robots here are being trained to learn, adapt and evolve... ultimately aiming to become helpful assistants or companions. [Stand_up] Wang Tianyu, Reporter: "Now the robot is following me, every move. And this is what you call learning. Clearly, I’m not a good teacher." [Voice_over] What these robots have to learn are the things we are often too lazy to do ourselves. For example, in factories, automated machines, which you can consider "non-humanoid robots," have already taken on a lot of the work that used to be done by human workers. [Sound_bite] Ruan Cheng, Test director, Embodied Intelligence Department, Agibot: "Those machines are for high-precision tasks, which are very rigid and require extensive programming. They cannot quickly adapt to changing environments." [Voice_over] Well, think about how many random variables are contained within making your everyday breakfast, cleaning up your kitchen and even sorting things out. Here are where the humanoid robots come into the picture. They need to repeat a single task about a thousand times to master each possible scenario … and there are countless scenarios to learn. [Sound_bite] Ruan Cheng, Test director, Embodied Intelligence Department, Agibot: "If you come here often, you'll notice that the scenarios change almost every week." [Voice_over] I had another question: Why do these robots have to be designed so similarly to a person? I mean: one head, two arms and a bit shorter than the average adult. [Sound_bite] Ruan Cheng, Test director, Embodied Intelligence Department, Agibot: "Our entire society is designed based on human ergonomics. A humanoid robot can make full use of the tools and infrastructure already created for humans, which makes it extremely convenient." [Voice_over] Having the ability to conduct working skills is not enough for robots to be intelligent, they need to perceive and understand the world and us. A 40-minute drive away from the training center, a company specializing in developing multimodal large language models, is using that to help robots process images, videos and text. [Sound_bite] Li Jing, Vice President, StepFun: "A robot's behavior follows two main steps: First, a highly generalized multimodal model interprets intentions, plans, tasks, and breaks them down—this is crucial. Second, the 'cerebellum' translates these plans into concrete movements." [Voice_over] So, the larger database a robot’s brain has, the broader knowledge it gets. Both knowledge and skills are important for humanoid robots to come into our home. [Sound_bite] Li Jing, Vice President, StepFun: "In the long run, the real breakthrough will be integrating multimodal models with embodied intelligence and applying them in real industries. But the day of true adoption may come soon." [Voice_over] By the end of last year, China has more than 450-thousand intelligent robot-related companies. For many of them, the next target that they want to achieve is probably to make the robots finish self-directed learning, or in other words, to teach themselves. [Restrictions : No access Chinese mainland]