Beyond PID

How can a quadcopter remain stable when unpredictable wind pushes it away from its target? Our project investigates whether modern deep reinforcement-learning controllers can outperform a traditional PID controller in this challenging control task.

We developed a GPU-accelerated simulation of a Crazyflie 2.0 quadcopter and trained six reinforcement-learning controllers, covering both on-policy and off-policy approaches. Every controller was evaluated under the same conditions: taking off, reaching a target position, and maintaining a stable hover in both calm air and persistent wind with intermittent gusts. Their performance was assessed using measures such as positioning accuracy, flight stability, control smoothness, success rate, and response time.

The exhibit presents our experimental process, simulation environment, controller designs, and key findings. Our accompanying poster explains how the algorithms were trained, tuned, and compared through a carefully controlled and reproducible benchmarking pipeline. Together, the two complementary studies by Wietske van der Westhuizen and Mathew Maartens explore the opportunities (and practical limitations) of using artificial intelligence for robust autonomous flight.