At WRC, a Real Robotic Café Serves Crowds and Clears Tables

Inside the booth, a mechanical arm behind the counter brewed pour-over coffee, and a human barista made latte art. A wheeled semi-humanoid robot, nicknamed Xiao K, carried cups through the crowd, avoided a spectator taking photos, recognized a camera as a personal item, and placed coffee on an empty table. After the customer left, the robot collected the empty cup and a used tissue, threw them into a trash bin, and sanitized its hands under an ultraviolet lamp. No paths were preset; every move was decided on site.

These tasks involved dynamic scene perception, real-time path replanning, semantic understanding to tell personal belongings from trash, manipulation of soft objects, long-horizon task sequencing, and coordination with humans and other robots. Outside the café, a small production line and a bead-stringing challenge showed multi-robot collaboration and precise handling, suggesting readiness for manufacturing.

Unbounded Dynamics was founded in Beijing in March 2025 and has completed hundreds of millions of dollars in financing this year. Its K15 robot passed the full EU industrial CE certification in July 2026, the first embodied intelligence robot to receive that certification. The company has announced 700 million yuan in commercial orders, with the first batch already shipped to Europe.

Founder and CEO Zhang Yufeng previously served as a vice president at Horizon Robotics, leading its smart driving business. Co-founder and CTO Xia Zhongpu, a PhD from the Chinese Academy of Sciences, ran prediction at Baidu Apollo and was in charge of end-to-end autonomous driving at Li Auto. Both experienced the shift from rule-based to end-to-end approaches in autonomous driving, which shaped their view on robotics.

Zhang was blunt about the popular VLA model: "VLA is essentially imitation learning. It doesn't hit a curb not because it understands what a curb is and the consequences of hitting it, but because it has seen many different scenes of curbs and data of avoiding them." He argued that data stacking cannot solve generalization in open environments. Unbounded Dynamics chose a latent-space world model, named MWA™, which focuses on physical changes and causal relationships instead of pixel-level prediction. The model topped the Stanford-linked RoboCasa GR1 TableTop benchmark with a 75.2% average success rate, ahead of Nvidia's GR00T-N1.6 and Xpeng's DIAL.

The company is pursuing both industrial and commercial applications. Industrial projects, such as safety-belt operations for ZF and battery factories for Envision, verify manipulation precision in controlled settings. The coffee shop tests continuous decision-making in an unpredictable environment. Both lines share the same MWA™ brain, and data from real use feeds back to improve the system. Over five days, the café served more than 1,000 visitors, with Xiao K working steadily while human staff tired. This dual-track approach provides a rare, verifiable evidence chain for embodied intelligence commercialization, from research results and benchmark scores to actual orders and operations.