WRC 2026's Only True 'Human-Machine Symbiosis' Booth Is an Immersive Robot Coffee Shop

The World Robot Conference (WRC) in Beijing this year is full of booths where robots repeat preset actions inside cordoned-off areas. The most unusual booth, however, is a fully operational robot coffee shop: a robotic arm makes pour-over coffee behind the counter while a human barista works on latte art, and a wheeled, semi-humanoid robot waiter wearing an apron and hat weaves through the crowd, dodges visitors taking photos, and places cups steadily on tables.

According to a report by QbitAI (量子位), the immersive robot coffee shop is not a one-off demo built for the show — the model had been trialed at a high-end café for some time and was only brought to WRC after stability and safety expectations were met. Over the five-day show, the booth drew more than 1,000 visitors, and the robot waiter, nicknamed Xiao K, kept running smoothly through real foot traffic, shifting table states, and changing lighting.

Behind the scene is embodied AI company Wujie Dongli (Boundless Power), founded in Beijing in March 2025 and backed by hundreds of millions of dollars in funding raised in the first half of this year. The company follows a full-stack in-house route covering brain, body, and core components; its K15 robot passed the full EU industrial CE certification in July 2026, becoming the first embodied intelligence robot worldwide to earn industrial-grade CE certification. Announced commercial orders total 700 million yuan, with the first batch of products already shipped to Europe.

At the booth, the robot must handle dynamic environment perception, real-time path replanning, semantic understanding (telling personal items from trash), soft-object manipulation such as tissues, and long-horizon task planning — serving, cleaning, collecting, and returning — while coordinating with other robots and humans. Each of these is an independent frontier research topic in robotics, yet Xiao K ran the entire chain continuously in an open space.

The company's technical bet is drawing particular attention. Founder and CEO Zhang Yufeng argues that VLA is essentially imitation learning, and that data stacking alone cannot solve generalization in open environments. Instead, Wujie Dongli chose a latent-space world model approach: its embodied general-purpose brain MWA™, released in June 2026, ranked first globally on the RoboCasa GR1 TableTop benchmark led by Stanford University with a 75.2% average task success rate, beating mainstream models including NVIDIA GR00T-N1.6 and XPeng DIAL.

At the WRC main forum, the World Model Expert Committee of the Chinese Institute of Electronics was officially established, with Zhang Yufeng invited as a committee member and joining the panel on the real distance from world models to serving humans. The founding team comes from autonomous driving: Zhang previously served as VP of Horizon Robotics and head of its intelligent vehicle business, while co-founder and CTO Xia Zhongpu led Baidu Apollo's prediction module and Li Auto's end-to-end efforts.

On commercialization, Wujie Dongli runs both industrial and commercial tracks: industrial deployments validate skills in fixed scenarios such as ZF seatbelt operations and Envision battery factories, while the coffee shop validates sustained decision-making in open environments. If embodied intelligence is ultimately headed for fully open settings like homes and outdoors, generalization must be tested and iterated in more uncertain environments first.

The coffee shop scene and the RoboCasa leaderboard are essentially two ways of validating the same logic — the ability to cope with change. The questions ahead are whether the latent-space world model route, which deviates from the VLA mainstream, can keep delivering under commercial and scale pressure, and when robots like Xiao K will truly step out of the expo hall into everyday service.