Gasgoo Munich- "In current embodied AI, the brain is ready; it is the hands that are holding things back." On July 3, at the 2026 Embodied AI Industry-Scenario Integration Conference hosted by Gasgoo Embodied AI, Cang Yu, vice president and general manager of the industrial division at Beijing Galbot Co., Ltd. (Galbot), identified the industry's most pressing bottleneck.Cang Yu, vice president and general manager of the industrial division at Beijing Galbot Co., Ltd. (Galbot)Beneath that lies a more fundamental dilemma. While peers race to build end effectors, they deploy one type of hand for Scenario A and swap in another for Scenario B. "It is hard not to conclude that this is just non-standard automation plus," Cang noted. He urged developers of core components like dexterous hands and tactile sensors to accelerate their efforts and deliver an end effector capable of both delicate manipulation and meaningful payload capacity.Cang drew a sharp contrast across three domains. GPT has already consumed three decades of internet data; autonomous driving has Tesla's FSD, with 7 million to 8 million drivers globally feeding it a steady stream of expert data. Chinese autonomous driving firms, meanwhile, invested heavily in buying vehicles and deploying fleet surveys early on—yet even they still lag behind FSD.For embodied AI, the data drought is far more severe. The automobile has kept essentially the same structural blueprint since 1886—an established tool of production where autonomous driving is merely an upgrade to the user experience. Robots, by contrast, emerged as physical forms only after large models revealed the potential to replace human labor. There is no legacy data to draw from, and certainly no factory willing to open up its proprietary process data."The chasm between simulated data and the real world persists," Cang acknowledged. "Even with the world's best simulations—even better than AMD's—we still face a gap of several percentage points." Capturing real-world data, he admitted, remains the most urgent problem to solve in the embodied AI space.Galbot's solution is to build demonstration production lines alongside partners in 3C manufacturing, auto manufacturing, and auto parts, sharing the rights to their process data. On the simulation front, the company has used purely simulated environments to train long-horizon, flexible-object tasks—like folding clothes—achieving a 100% success rate.As for the pace of commercialization, Cang was unequivocal: "The most pressing issue for us and our peers right now is simply solving the problem—efficiency comes later." He urged the industry to be patient.In industrial settings, Galbot's strategy treats the foundation model as a "high school diploma" for the robot. The industrial division then applies vertical-specific data to shift the model from broad generalization to narrow specialization. "Industry demands limited generalization, strict cycle times, and high reliability," Cang explained. "I do not need a robot that can do everything—I just need it to perform under fixed procedures."On the cost side, Galbot deployed its first pharmacy robot in December 2024 and has since scaled to nearly 100 locations—making it "the first scenario where we have achieved ROI break-even." In industrial logistics, its transfer robots are operating with nearly 100 units at CATL, with additional deployments at Foxconn and Siemens.Finally, Cang called on industry associations to drive standardization. "It is just like the early days of smart factories—data silos meant a single plant might run several MES platforms, and that problem persists today." Without unified standards, embodied AI risks "falling into prolonged fragmentation." He added: "Whether it is the broader industry or our direct peers, everyone needs to stay open-minded—and be willing to remain patient during periods of slow progress. This is a long-haul industry."