Reinforcement Learning for UAV Landing on Aggressively Moving Ground Vehicles

초록

This study proposes a reinforcement learning (RL) model for autonomous landing of an unmanned aerial vehicle on an aggressively maneuvering ground vehicle (GV), where the GV acceleration exceeds 1.5 m/s2. The key challenges are maintaining reliable onboard visual perception during landing and determining the appropriate time to touchdown under unpredictable GV motion. To address these challenges, this study proposes a kinetic-energy-based touchdown safety penalty that provides explicit feedback on landing safety and discourages unsafe touchdown. In addition, a continuous perception quality function is incorporated to evaluate the detection reliability of AprilTag and promote perception-aware behaviors. In simulations over 300 aggressive trajectories, the proposed method outperformed tuned baselines by achieving an 89.0% landing success rate and a 0.7% tag-detection failure rate. Real-world experiments across multiple GV motion scenarios further validated the proposed model, including stable landings at the GV driving up to 5.13 m/s and acceleration up to 2.87 m/s2. These results indicate that explicitly optimizing observability and touchdown safety leads to reliable landing decisions under aggressive GV maneuvers.

키워드