Learning Agile Flight in the Wild

Learning Agile Flight in the Wild, published in Science Robotics 2023, represents a major breakthrough in autonomous drone racing and acrobatic flight. The key paradigm shift is that the entire policy is trained end-to-end in simulation using deep reinforcement learning, yet transfers to the real world with zero additional fine-tuning on real data. This is achieved through extreme domain randomization over visual appearance, dynamics, and sensor noise, paired with a compact network architecture (ResNet-18 + LSTM) running at 20 Hz onboard an NVIDIA Jetson. The authors formalize the problem as a partially observable Markov decision process where the drone must track a state trajectory (roll, pitch, yaw rates, and velocity) while avoiding crashes. The policy directly outputs low-level actuator commands from raw camera images and IMU data, bypassing any explicit state estimation or motion planning pipeline. During training, the simulation randomizes camera pose, exposure, lighting, background textures, wind gusts, mass distribution, and motor delays — over 50 parameters in total. This pushes the policy to learn robust visual-inertial features that are invariant to the exact simulation environment. Real-world validation is conducted both on an autonomous racing drone and in freestyle acrobatic sequences including loops, rolls, and power loops. The trained policy achieves lap times competitive with professional human pilots on a physical racing track, and acrobatic maneuvers that match or exceed the best hand-designed controllers. Ablations show that visual randomization is the most important factor for sim-to-real transfer, followed by dynamics randomization. The work also includes open-source training code and a realistic simulator built on top of Flightmare. For UAV researchers and practitioners, this paper is a masterclass in sim-to-real transfer. It demonstrates that with sufficient diversity and a compact policy architecture, sim-only training can reach world-class real-world performance. The methodology has since been adopted widely in drone control, autonomous racing, and aerial cinematography research.