Brief

Researchers have developed PerchRL, a reinforcement learning framework designed for quadrotors to autonomously perch on moving, inclined platforms. This system utilizes a two-stage learning process, starting with state-based pre-training and then fine-tuning with visual input. To enhance its ability to handle unpredictable platform movements and intermittent visual loss, PerchRL incorporates randomized trajectories, temporal augmentation, and visibility-aware state augmentation with active perception rewards. Both simulations and real-world tests have confirmed the system's stability, real-time performance, and adaptability across different quadrotor models. AI