X Square Robot, a Beijing-based embodied-AI startup, has open-sourced HOST, an inference-time learning framework that lets a humanoid robot watch a 29-second human demonstration and reproduce the skill at 62 percent success. The approach flips the embodied-AI recipe from offline fine-tuning to on-the-fly imitation. The industry standard requires roughly 50 expert demonstrations and four hours of fine-tuning to clear 38 percent. HOST inverts the recipe by asking the robot to imagine the end state and reverse-engineer the motion, using a vision module, an action module, and a dynamic time-warping alignment module that drops alignment error by roughly an order of magnitude. The framework is rooted in the cognitive-science theory of simulation, replacing offline fine-tuning with inference-time in-context learning, where the robot consumes the video as a prompt. X Square Robot was founded in late 2023 and has raised more than one billion yuan across eight funding rounds, with Meituan Strategic Investment, Alibaba Cloud, Lenovo Capital and Delian Capital among the backers. The open-source release is a deliberate bet on developer mindshare as a route to enterprise pilots. The 62 percent number is a research benchmark, not a deployment-ready home robot, but it resets the conversation about what embodied-AI training pipelines will look like once inference-time learning becomes the default.