Vision, flies, neurons and robots

Summary form only given. From the knowledge we gained about the processing of motion in the fly visual system we explored how an agent can use motion vision to control its locomotion. We built a biologically inspired visual system made of opto-electronic motion detecting "neurons". Instead of involving a von Neumann architecture, this system is based on parallel and analog networks, as is the case with nervous systems in general. This visual system now guides a mobile robot without collision in unforeseen environments. Computation taking place onboard our robot essentially relies on brainlike, parallel, analogue, continuous time, asynchronous networks, with many additional features common to advanced animal visual systems, such as nonuniform retinal sampling, retinotopic projections of sensory maps, saccadic suppression, and corollary discharges that inhibit vision during eye rotations. All circuits are hardwired, yet adaptive behavior is achieved in the sense that the robot immediately copes with novel environments.