Safely Learn to Fly Aircraft From Human: An Offline–Online Reinforcement Learning Strategy and Its Application to Aircraft Stall Recovery

Researchers have made many attempts to apply reinforcement learning (RL) to learn to fly aircraft in recent years. However, existing RL strategies are usually not safe (e.g., can lead to crash) in the initial stage of the training of an RL-based policy. For increasingly complex piloting tasks whose representative models are hard to establish, it is not safe to apply the existing RL strategies to learn an RL-based policy by interacting with an aircraft. To enhance the safety and feasibility of applying an RL-based policy to an aircraft, this study develops an offline–online RL strategy. The offline–online RL strategy learns an effective initialization for an RL-based flight control policy from human pilots without interacting with an aircraft through offline RL. Then, the offline–online RL strategy can further improve the RL-based flight control policy safely without leading to crash by interacting with the aircraft according to regular online RL, requiring no or very little intervention performed by a human pilot. To demonstrate and evaluate the offline–online RL strategy, the strategy is utilized to address the stall recovery problem of aircraft based on a flight simulator.