Array- and pixel-level strategies for low-data and low-power vision in embodied computing image sensors

Machine vision systems rely on visual information from the surrounding environment to support task-level decisions, such as navigation and object interaction. Image sensors serve as the front-end interface of these systems, where visual data are first acquired. In machine vision frameworks, image sensors function as passive data acquisition units, while most perceptual processing occurs in external computing hardware, creating a data-transfer bottleneck that increases computational burden. Embodied computing image sensors address this limitation by reconfiguring the sensor architecture to enable task-relevant visual processing during image acquisition. These strategies reduce computational load by shifting part of the processing to the sensor itself, enabling low-data and low-power vision. In this perspective, we discuss these developments and outline directions toward unified sensing platforms for computation-efficient vision systems.