Vision-Guided Outdoor Flight and Obstacle Evasion via Reinforcement Learning
Researchers have developed a novel reinforcement learning policy for drones to autonomously navigate and evade obstacles in unknown outdoor environments. The system utilizes stereo-vision depth and visual-inertial odometry to generate velocity commands for commercial drones. Trained in simulation with a two-stage process and fine-tuned with domain randomization, the policy demonstrated successful zero-shot transfer to new drone platforms and environments without prior exposure. AI
IMPACT Enables autonomous drone operation in GPS-denied and unknown environments, potentially for delivery or surveillance.