Nano quadcopters are small, agile, and cheap platforms well suited for deployment in narrow, cluttered environments. Due to their limited payload, nano quadcopters are highly constrained in processing power, rendering conventional vision-based methods for autonomous navigation incompatible. Recent machine learning developments promise high-performance perception at low latency, while novel ultra-low power microcontrollers augment the visual processing power of nano quadcopters. In this work, we present NanoFlowNet, an optical flow CNN that, based on the semantic segmentation architecture STDC-Seg, achieves real-time dense optical flow estimation on edge hardware. We use motion boundary ground truth to guide the learning of optical flow, improving performance with zero impact on latency. Validation on MPI-Sintel shows the high performance of the proposed method given its constrained architecture. We implement the CNN on the ultra-low power GAP8 microcontroller and demonstrate it in an obstacle avoidance application on a 34 g Bitcraze Crazyflie nano quadcopter.

Flying insects are capable of autonomous vision-based navigation in cluttered environments, reliably avoiding objects through fast and agile manoeuvres. Meanwhile, insect-scale micro air vehicles still lag far behind their biological counterparts, displaying inferior performance at a fraction of the energy efficiency. In light of this, it is in our interest to try and mimic flying insects in terms of their vision-based navigation capabilities, and consequently apply gained knowledge to a manoeuvre of relevance. This thesis does so through evolving spiking neural networks for controlling divergence-based landings of micro air vehicles, while minimising the network's spike rate. We demonstrate vision-based neuromorphic control for a real-world, continuous problem, as well as the feasibility of extending this controller to one that is end-to-end-learnt, and can work with an event-based camera. Furthermore, we provide insight into the resources required for successfully solving the problem of divergence-based landing, showing that high-resolution control can be learnt with only a single spiking neuron. Finally, we look at evolving only a subset of the spiking neural network's available hyperparameters, suggesting that the best results are obtained when all parameters are affected by the learning process.