Ant Lingbo CEO Says Robot Deployment Still Held Back by Capability and Cost

At the 2026 Bund Summit, Zhu Xing, CEO of Ant Lingbo, told APPSO that large-scale robot deployment still faces capability and cost bottlenecks, even for everyday scenarios such as convenience stores. He said the overall model ecosystem is still in a cold-start phase, and most deployable products rely on imitation learning, which struggles with unexpected situations and long-horizon tasks.

Unlike many robot makers focused on complete hardware, Lingbo concentrates on embodied intelligence "brains". Zhu introduced that the company is training LingBot-World, an interactive world model, LingBot-Video, an embodied video generation foundation model, and LingBot-VLA 2.0, an action model trained on top of it. The aim is to have one foundation model adapt to different robot embodiments and tasks, with post-training lowering the barrier for specific applications. The solution has been deployed at Guoda Pharmacy retail stores, with demonstrations covering drug picking, logistics sorting, and industrial loading and unloading.

Zhu believes the value of foundation models lies in improving few-shot performance in post-training: achieving the same results with less data, or better results with the same data. But scaling across stores remains hard due to generalization gaps and differing environments. He also said today's robots cannot yet digest "rough" everyday data, and that data quality and distribution matter more than raw volume. He predicted that productivity-focused applications could grow quickly from the second half of 2026 through 2027-2028, and he expects two exponential changes in 2027: a rise in the number of real-world robots driven mainly by models, and a sharp drop in post-training and data costs. On the prospect of a "ChatGPT moment" for embodied intelligence, he said the "Chat" part matters more than "GPT": applications need to reach a certain scale to close the loop.