Cooperative pursuit-evasion control
Pursuit-evasion scenarios are a prevalent problem in the real-world, for example, autonomous search and rescue, aerospace defence and remote surveillance.
A pursuit-evasion problem is considered in which a team of slower, less agile pursuers must collaborate to capture a highly agile, high speed evader. The current leading approach only captures evaders exhibiting specific behaviours.
Deep reinforcement learning is an emerging area for multi-agent control which has demonstrated unprecedented success in complex multi-agent tasks such as games like Starcraft 2 and Capture the Flag. Using a custom deep reinforcement learning architecture, we demonstrate that this new technique consistently outperforms the benchmark. The cooperative teaming methods were loaded into hardware, using a team of drones to demonstrate their effectiveness. A first of its kind drone control system was used, which improves the response between the autonomous commands and drone behaviour.