Work

Neuro-evolution learned neuromorphic control for a vision-based 3D landing

Year: 2023
Type: dissertation
Abstract: Neuromorphic control is a biologically inspired processing method that uses Spiking Neural Networks (SNN) to simulate the computational principles observed in mammal nervous systems.Micro Air Vehicles (MAV) are small and light air vehicle systems of which the level of intelligence is dictated by sensors, algorithmic complexity and onboard computing power.Implementing flight control systems with SNNs would allow us to pursue an end-to-end event-based processing sytem that decreases scale and power usage, which can substantially increase the artificial intelligence of MAVs.With the rise of available neuromorphic sensors and processors, one would expect to find many robotic applications.However, learning and the reality gap impede robotic applications.This work shows that a single SNN controller can be evolved for 3D flight control (thrust, roll and yaw).By learning an Optical Flow (OF) based flight control strategy, it is proven that a neuromorphically controlled MAV can take-off, cruise and land on a target with great precision (π‘₯ πœ– , 𝑦 πœ– < 0.02π‘š & 𝑉 𝑧 < 0.05π‘šπ‘  -1 ).The extracted ventral flow in the longitudinal axis controls motion in the for π‘₯ and 𝑧 in the π‘₯𝑧 plane.By means of evolution, not only are the SNN network parameters evolved, the topology is modified during training as well.Instead of using a pre-defined fully connected Multi Layer Perceptron (MLP) like architecture, a randomly connected and sparse network is custom built and learned.
Author: Lodder, Erwin (author)
Language: English
Open Access status: green
OpenAlex ID: W7058466197
Cites: 144
Cited by:
Related to: 10