A Custom Fabricated Low Weight On-Board Vision Sensor for Insect Scale Robot

Abstract

Controlled flight of insect scale ( 100 mg) Micro Aerial Vehicles (MAVs) has to date required offboard sensors and computation. Achieving autonomy in more general environment will require integrating sensors such as a camera onboard, but this is a challenging task because of the small scale as the component mass and computation must be minimized. In this work we present the design and fabrication of a low-weight camera 26 mg mounted on a flapping wing insect scale aerial and ground robot. We trained a Convolution Neural Network (CNN) with the images captured by the camera to classify flower and predator images. We show that feedback from the CNNs classification can command the robot to move toward flower and away from predator images. Our results indicate that these computations can be performed using low-weight microcontrollers compatible with the payload and power constraints of insect-scale MAVs. We also perform preliminary optic flow based position estimation experiments with the low weight camera. Many desired capabilities for aerial vehicles, such as landing site selection and obstacle detection and avoidance, are ill-defined. This work shows that Computer Vision (CV) and CNNs, which have previously been deployed only on larger robots, can now be used on insectscale for such tasks.