咖啡空间成具身大脑训练场？WRC世界模型热度高涨：技术路线仍未收敛，能力徘徊于“学步”与“青春期”之间Is coffee space a training ground for embodied brain? The WRC world model is gaining popularity: the technological roadmap has not yet converged, and the ability is hovering between "learning to walk" and "adolescence"

The 2026 World Robot Conference (WRC) kicked off on August 19th in Yizhuang, Beijing. In this year's WRC exhibition hall, there is a special coffee space: there are human baristas inside the counter, long-range hand drawn robotic arms on the counter to assist in making coffee, and wheeled robots outside the counter shuttle through the crowd to provide services to customers.

Unlike the 24-hour intelligent unmanned coffee shop under JD's Seven Fresh Coffee, which has just landed in Beijing Galaxy SOHO, the coffee space in the WRC exhibition hall is more focused on "human-machine symbiosis". This is the training ground selected by Wujie Power, a embodied intelligence company that has only been established for over a year and completed hundreds of millions of dollars in financing in the first half of this year, for its embodied brain.

Why is it a coffee space?

Zhang Yufeng, founder and CEO of Wujie Power, answered pragmatically during an interview with media including Daily Economic News reporters at the exhibition hall: the tasks and SKUs of items in the coffee shop scene are relatively concentrated, unlike the vast differences in the home environment - each family's bed sheets, clothes, and tea cups are different. But coffee shops also have enough background generalization: the lighting is changing, the flow of people is changing, and the personal items left by customers on the table are changing. Robots must recognize and avoid them, and when collecting tissues, they cannot accidentally collect mobile phones. When placing coffee, they must choose a suitable location.

In Zhang Yufeng's view, this scenario has clear business boundaries and dynamic environmental variables, which allows the company's self-developed hidden space world model to continuously learn within a stable business framework, while also having to deal with massive real changes, thus testing its ability to transfer existing skills to different personnel, locations, and states.

Every journalist has noticed that world models that can understand and predict the physical world are now the emerging technological focus of the industry and one of the hot topics of this year's WRC. But the current industry has not yet entered the stage of "technological convergence", and behind the coffee space is the competition of various technologies and commercialization routes.

Coffee scene becomes a popular experimental field for robot landing

Entering the exhibition hall of this year's WRC, the most intuitive feeling is that robots are no longer just performing on stage or fighting on the boxing ring, but are showcasing their true abilities to work in various scenes. Among them, "making coffee" has become a popular "test field" for testing the real landing ability of robots.

Shenzhen Anno Robotics is the only company in the beverage robot segment that has won an independent booth, and also appeared at the booth of its partner JD.com. Every time the reporter saw it on site, their dual arm embodied robot latte art coffee printing bar and single arm AI latte art coffee printing booth could quickly complete a cup of coffee with exquisite latte art through high-precision robotic arms, visual recognition, and machine learning. Users could also upload AIGC (Artificial Intelligence Generated Content) images to select unique printing patterns.

In the "Robot Cafe" jointly presented by Qinglang and Novacoffee, the humanoid robot XMAN-R1 appeared as a "specially appointed barista". The robot is based on the KEENON ProS, a specialized vertical domain model trained on the self-developed general VLA (vision language action) architecture by Qinglang. XMAN-R1 can independently complete the steps of cup picking, coffee extraction, etc., and place the prepared coffee on the latte machine for customers to DIY print.

Different from the whole process unmanned, closed vending machine coffee robot featured by JD Seven Fresh Coffee, there are still human baristas in the coffee space of boundless power, and the robot does not stay behind the closed space or the bar, but has become a waiter in the coffee space, taking the coffee end from the real baristas to customers, collecting empty cups, cleaning the table, and even taking the initiative to do ultraviolet disinfection and sterilization for both hands after throwing garbage.

In commercial service scenarios, customers usually expect to see people, and few coffee shop brands want to become completely unmanned coffee shops, so human-machine symbiosis is the core. "Zhang Yufeng emphasized," We no longer surround robots, but let them truly cooperate with people

The "Human Machine Symbiosis Immersive Robot Coffee Space" in the Boundless Power Exhibition Area. Photo source: Photo by journalist Zheng Xinwei

Every time the reporter noticed on site, the real barista was wearing the AnySense EGO first view multimodal data acquisition device developed by Boundless Power on their head, which was used to collect real operation data and provide feedback for subsequent model upgrades. In Zhang Yufeng's view, the iteration of hard technology needs to be deeply integrated with scenarios and B-end customers, finding opportunities to lay eggs along the way, using real scenario data to drive high-quality data feedback, forming a data flywheel, and promoting continuous iteration of embodied brains.

According to the reporter, in terms of commercial service scenarios, Wujie Power has entered the coffee shop market. In the next four months, it will continue to cooperate with a coffee shop brand in Beijing for trial operation and gradually transition to normal operation. It plans to enter the Korean market in the fourth quarter of this year; After the coffee shop, fast food restaurants and hotels are the next targets. In the industrial scene, boundaryless robots have cooperated with automotive parts companies and energy companies to truly enter the production line.

For example, for origami boxes, foreign-funded car factories in Beijing have demands for robots to help fold boxes and organize parts. We have created robots that can fold paper boxes in industrial and household scenarios, achieving 'one machine, multiple functions'. In most exhibitions, robots may only do one thing, but our stacking socks and clothes is actually a model. If we throw the items to it, it will handle them one by one, "said Zhang Yufeng.

At the exhibition site, Boundless Dynamics also showed that three robots based on the same model realized independent division and collaboration around folding, classification, packaging and other links. The lack of generalization ability is also the biggest bottleneck that restricts robots from entering daily life and households, as mentioned by Wang Xingxing, the founder of Yushu Technology, in his speech at the conference. "Generalization" is also a key word repeatedly mentioned by Zhang Yufeng. The generalization ability of "the first mock examination with multiple abilities" is the direction that he thinks the embodied brain needs to work hard.

The world model technology roadmap has not yet converged

Driven by the World Labs founded by AI godmother Li Feifei, the highly popular world model route presented even greater momentum at this year's WRC: the Beijing Humanoid Robot Innovation Center released the embodied multimodal unified model PelicanUnify, Daxiao Robot brought the enlightened world model, and the excellent vision showcased the new generation of embodied basic model GigaBrain-0.7... Exhibitors who bet on the world model route showcased their embodied brains at this year's WRC.

At the Super Dimensional Power booth, a full-size humanoid robot KAIBot engaged in a heated table tennis match with the live audience. Faced with a fast and unpredictable ball, the robot completes recognition, judgment, and hitting within a millisecond window. Supporting this capability is its self-developed KAI world model. The model is built around a three-step closed-loop of "generating a virtual world for action interaction - model evaluation and understanding - mastering the laws of the physical world", allowing the robot to complete massive pre training in the simulation world and understand the operation of ping-pong matches before entering real confrontation.

In his speech at this year's WRC main forum, Wang Xingxing, founder of Yushu Technology, also revealed that Yushu Technology had been exploring the direction of world models based on video generation as early as the beginning of 2020. However, due to unsatisfactory results, it was put on hold until it resumed investment last year. "Our company has always invested heavily in AI models, and it should be the direction with the largest capital and manpower investment at present.

The unbounded power that also bet on the world model chooses the "hidden space world model+reinforcement learning" technology route, and Zhang Yufeng does not hide his confidence in this route, even using the phrase "closest to the end".

This entrepreneur, who has been deeply involved in the automotive industry for many years and has served as the Vice President of Horizon Robotics, used autonomous driving as an example to compare the popular VLA (Visual Language Action) route: "The car does not press the curb because it has seen millions or millions of trajectories without pressing the curb, rather than truly understanding that it is a curb and the consequences of pressing it like a human." Zhang Yufeng pointed out that VLA is essentially imitation learning, rather than a true understanding of spatiotemporal correlations and physical world causality.

Regarding the debate between VLA and world models, Zhang Yufeng candidly stated: "Today, there is no ultimate path for physical AI or AGI (General Artificial Intelligence), and autonomous driving has not reached the ultimate AGI path. Otherwise, why hasn't L4 scale emerged? ”He expressed his hope to see greater innovation in algorithms and models within the industry. He also revealed that Wujie Power will release a new version of the base model in September with larger parameters and without relying on any existing open source models.

In addition, Zhang Yufeng also pointed out that algorithms, hardware, and scenarios are the three in one iteration of the robotics industry, but the hardware maturity and standardization of the entire industry chain still need to be improved. At present, the market size of humanoid robots is still small, which poses challenges to hardware maturity polishing, cost reduction, and consistency improvement. The Chinese supply chain is indeed the strongest, but it does not mean it is completely mature. We are very strong in 'having and not having', but in terms of cost and consistency, the entire industry still has a long way to go

Regarding the current development stage of the embodied intelligence industry, Zhang Yufeng judged that "the current ability of embodied intelligence is probably between the toddler stage and adolescence, able to do some basic work, but too complex things cannot be learned at once." However, he likes the theme repeatedly mentioned in the opening video of this conference - "Try again.

Robots have gone from running in the past to working in factories now, and there will always be times of failure, just like how humans grow from childhood to adulthood. Trying again will lead to success. "In Zhang Yufeng's vision, embodied intelligence will eventually become a new type of infrastructure, like new water, electricity, coal, and new computing infrastructure, from realizing current value creation to collecting more data, fuel, and feedback, forming stronger capabilities and becoming a new type of infrastructure.

Disclaimer: The content and data in this article are for reference only and do not constitute investment advice. Please verify before use. Based on this operation, the risk is borne by oneself.

Daily Economic News