Colloquium: Reinier Vos (C&O)
04 October 2024 10:00 - Location: Lecture Hall H, Faculty of Aerospace Engineering, Kluyverweg 1, DELFT | Add to my calendar
Hunt like a fly, strike like a drone: learning pursuit controllers for insect interception through multi-agent deep reinforcement learning for onboard use in autonomous quadcopters
Insect pest elimination through MAV interception can reduce the need for insecticide and contribute to sustainable agriculture. In this research, we analyze the feasibility of such solutions through simulated two-player differential games of pursuit and evasion with agents operating on minimalistic sets of biologically-plausible observations and optimized through deep multi-agent reinforcement learning. Our pursuer and evader agent, representing the quadcopter drone and insect pest respectively, are asymmetric in design, capabilities and objectives. From our results, we show that our quadcopter pursuer is consistently able to pursue and intercept an agile and reactionary insect-inspired evader as well recordings of actual insect targets. We remark that our pursuer implements motion camouflage, drawing comparison to the hunting strategy of dragonfly. Conversely, we do not find decisive evidence for the need of a reactionary evader and multi-agent optimization. We attribute general interception efficacy to the pursuit controller’s ability to set appropriate references conditional on drone and target state, in consideration of the sluggish vehicle model at hand.
Supervisor: M. Yedutenko