Benedict was working on "Simulation and Reinforcement Learning for Micro-Scale Cargo Transport" supervised by Samuel Tovey. We congratulate him on this wonderful achievement and invite you to attend the defense.
Abstract: Microrobots a few micrometres across push with only tens of femtonewtons, so moving anything larger than themselves takes a swarm, working against thermal noise that scrambles their heading within minutes. Reinforcement learning can train such swarms to transport cargo or rotate a rod, but the trained network does not reveal what rule it has learnt or how it behaves when conditions change. In this thesis I train swarms of simulated microrobots, calibrated to laser-driven particles, and then take the learned policies apart: I redeploy them across a range of temperatures, degrade their perception, and distil them into closed-form rules by symbolic regression. A cell-repair task extends the approach to learned communication. The learned strategies turn out to be simple rules in one or two variables that hold across the trained regime.
Best of luck Benedict on future endeavors!
