A data-driven approach for magnetic microswarm steering

One way to improve current medical standards is to investigate methods that aim to reduce the scale and severity of trauma caused while treating ailment. Here microswarms, collections of nano-scale particles that respond to electromagnetic forces, come into their own. Microswarms are a remotely controlled entity, requiring no onboard power or intelligence, as user controlled external electromagnets are used for articulation. Accurate simulation is key to improving this technique, and there exist microswarm simulators that can perform very accurate modelling. However, these simulations are not performed in real time, preventing implementation of human-in-the-loop control schemes. Human-in-the-loop control may result in improved patient trust and can aid in increasing input (and, thus, control) accuracy for users who are experts in their medical field, but not in the technology being used. This thesis documents a novel approach to fill the outlined gap, by implementing a human-in-the-loop control scheme into a bespoke real-time simulator, with haptic assistance to untrained users provided by a machine learning model. The efficacy of this system in its current form is compared to how a user may interact directly with the produced simulator. The real-time simulation was implemented in MATLAB and was verified against real world experimental data. Users provided input using a Novint Falcon haptic device. Data collection was performed by means of experimentation with participants. This data was then used as a training set for a neural network. To guide the microswarm to the goal outlet, the network was trained to map the spatial position of the particles in the swarm at each time step of the simulation to a suitable magnetic force to be applied. The neural network was implemented into a shared control approach interfacing with the haptic device. This form of haptic assistance effectiveness was evaluated in a further experiment with human participants.