In the realm of physical AI, edge devices grapple with constraints in power consumption and computational capability, posing a significant bottleneck to the development of real-time visual perception. Traditional visual perception chains are fraught with complex processes and substantial energy consumption, primarily due to the redundant transfer of data. The research team has ingeniously proposed shifting the token generation process to the sensor level and has developed the groundbreaking 'Opto-Semantic Core' chip. This chip comprises a photosensitive memory array and peripheral addressing circuits, with pixels that boast a trifecta of functions: photosensitivity, memory, and analog computing, facilitating the direct output of tokens. This breakthrough overturns traditional serial models, offering a novel approach to surmount the power and computational limitations of edge AI in the post-Moore era. It holds immense potential for applications across various domains, such as embodied intelligence, driving the widespread adoption of physical AI.