Trajectory Tracking with Zero-Shot Reinforcement Learning on Aerial Robots: A Real-World Evaluation

Reinforcement learning (RL) has demonstrated the ability to perform precise control in aerial robot applications. However, RL policies have struggled with the sim-real gap, which often results from the problem of incompatible modeling of environments. To address this problem, this work evaluates an RL policy that was developed for an aerial robot in real-world experiments with the trajectory tracking task. Our experimental results show that the RL policy generated only a 0.0327 m trajectory tracking error in the lemniscate trajectory using a zero-shot approach without requiring any tuning. Furthermore, this study validated the SimpleFlight framework for reproducibility with the different studies.