A newly emerging player in embodied intelligence has captured market attention. “Junao Panshi,” a startup founded in late 2025, has announced the completion of a new funding round worth hundreds of millions of yuan, although the exact amount and investor lineup remain undisclosed. The company stated that the round was led by top-tier industrial capital with deep expertise in brain-inspired computing and embodied intelligence, with participation from existing shareholders and several prominent funds.
Despite being less than a year old, the company has moved at an unusually rapid pace. Within just two months of its founding, it had already secured tens of millions of yuan in seed funding. What has further intrigued observers is not the scale of its financing, but its unconventional technological direction. While many industry entrants have gravitated toward the increasingly popular Vision-Language-Action (VLA) paradigm, Junao Panshi has instead chosen to rebuild robotic cognition through brain-inspired intelligence, focusing on the development, engineering deployment, and real-world validation of a Cognitive World Model.
This differentiated approach appears to have resonated with investors. Rather than emphasizing headline funding figures or high-profile backers, the company’s clearly defined commercialization pathway and distinct technical thesis have been central to its appeal.
The debate over technological direction in embodied intelligence has intensified in recent months. At the Sequoia AI Ascent conference, Jim Fan, a leading figure in the field, questioned the dominant trajectory of robotics development. He argued that current VLA systems are, in practice, heavily skewed toward language processing—effectively functioning as “LVA” models—resulting in insufficient encoding of physical reasoning and unreliable action execution in complex environments.
As embodied AI systems are deployed in increasingly dynamic real-world settings, the limitations of VLA models have become more apparent. High-quality data cannot be scaled indefinitely, computational resources remain finite, and models often struggle with generalization across environments, requiring retraining for new scenarios. Moreover, robots still lack persistent memory and the ability to learn continuously over time.
Fan suggested that within the next one to two years, the primary source of training data for robots will shift from costly human teleoperation to widely available first-person human video data on the internet. This shift implies a fundamental transition: robots must move beyond imitation and toward genuine understanding of the physical world.
This perspective is gaining broader consensus among industry leaders. Wang Xingxing, CEO of Unitree Robotics, has similarly stated that world models represent the ultimate solution for embodied intelligence, with VLA serving only as an interim approach. Yann LeCun, Turing Award laureate and one of the pioneers of deep learning, has also emphasized that world models will define the next generation of AI. His proposed Joint Embedding Predictive Architecture (JEPA) focuses on reasoning, planning, and modeling the real world, further reinforcing this direction.
In parallel, recent developments such as the LDA-1B model—jointly released by Galaxy General, NVIDIA, Tsinghua University, and Peking University—signal a broader shift away from reflexive imitation toward world-model-based learning.
Junao Panshi’s strategy aligns squarely with this emerging paradigm. From its inception, the company has committed to building a Cognitive World Model tailored to real-world physical environments. In its view, embodied intelligence is transitioning from “action intelligence” to “cognitive intelligence.” The next phase will not merely involve enabling robots to perceive tasks and execute actions, but equipping them with human-like capabilities: few-shot abstraction, multi-dimensional environmental perception, long-term memory, and active reasoning—allowing stable, cross-scenario performance in complex real-world conditions.