Active Perception

A robot deliberately moves its sensors or body to see or feel a target better, rather than just passively observing.

Active perception means an agent doesn’t just passively receive sensor data, but actively decides where, how, and when to look based on the task and its current judgment. The idea was proposed by Ruzena Bajcsy in 1985 and published in Proceedings of the IEEE in 1988; that same year Aloimonos and colleagues independently proposed “active vision.” A 2016 survey by Bajcsy, Aloimonos, and Tsotsos defined it this way: an agent knows why it needs to perceive something, chooses what to perceive, and decides how, when, and where to perceive it. It addresses the problem of a single viewpoint not giving enough information — occlusion, a target outside the field of view, or poor lighting. Common forms in embodied AI include turning the head for a different view, adjusting a wrist camera, or reaching out to touch something to confirm its material; it is closely related to next-best-view planning, interactive perception, and active exploration.

ExampleA cup is half-hidden behind a cardboard box; a humanoid robot turns its head and bends down for another shot from a different angle, confirms which way the handle is facing, and only then reaches to grasp it.

Also called

Active Vision, Active Sensing

Related

Interactive Perception · Next-Best-View Planning · Active Exploration · Perception-Action Loop · Embodied Perception · Head Camera

Sources

Revisiting Active Perception (Bajcsy, Aloimonos, Tsotsos, arXiv:1603.02729)