An Insect-Sized Robot That Uses a Custom-Built Onboard Camera and a Neural Network to Classify and Respond to Visual Input

To date, controlled flight of very small, insect-sized (rvlOO mg) Micro Aerial Vehicles (MAVs) has required off-board sensors and computation. Achieving autonomy in more general environments that do not have such resources available will require integrating these components. In this work we present advances toward this goal by demonstrating a new, custom-built, low-weight (26 mg) camera mounted on a 74 mg flapping-wing robot. We implemented a convolution neural network (CNN) to classify images. Using this system, we demonstrated how the camera-equipped robot could repeatably move toward flower images and away from predator images. An estimate of the computational requirements of our network indicates that it could be performed using low-weight microcontrollers that are compatible with the payload and power constraints of insect-scale MAVs. Many desired capabilities for aerial vehicles, such as landing site selection and obstacle detection and avoidance, are ill-defined because the boundary between positive and negative classes are unclear. This work shows that CNNs operating on input from vision, which have previously been deployed only on larger robots, can be used at insect-scale for such tasks.