In a surprising development, New York University’s Agile Robotics and Perception Lab has streamlined the process of training drones to achieve basic autonomy. Using a curriculum-driven sample efficiency technique, the lab’s system can train a drone in simulation to fly in a stable and controllable manner within just 18 seconds. This approach, which leverages deep-reinforcement learning algorithms and Apple’s M-series processors, can be applied to drones of various sizes and configurations. The open-source system is available for anyone to use and further improvements are expected, including integration with the PX4 open source drone autopilot. The research paper and necessary resources can be found on GitHub.
https://spectrum.ieee.org/drone-quadrotor