Efficient neuromorphic optomotor heading regulation

“Neuromorphic” vision sensors are a recent development in sensing technology. They can be thought of a camera sensor whose output is a sequence of “retinal events” rather than frames. Events are generated independently by each pixel as they detect a change in the light field. These sensors have low latency (<;10 microseconds), high dynamic range (>120 dB), and very low power consumption. Therefore, they are well suited for control applications where power is limited yet high performance is necessary. Existing computer vision algorithms that work on frames cannot be adapted to process retinal events from neuromorphic sensors, so a new class of algorithms needs to be investigated. This papers considers the problem of designing a regulator for the heading of a vehicle based on the feedback from an on-board neuromorphic sensor. It is shown that a nonlinear function of the events retinal positions, followed by retinal integration, followed by a linear filter is a simple design that is sufficient to guarantee stability. This shows that computationally simple controllers are sufficient to control motion tasks even with the feedback from noisy and ambiguous event data, and without having to compute explicit representations for the state.