Drone Altitude Control with Reinforcement Learning

In this paper, we build a simulated flight environment based on PyBullet continuous control tasks. Soft Actor-Critic (SAC) algorithm was selected as the final algorithm after testing and comparison. The model was trained to hover in the air as the target, and the reward in the training process was compared to illustrate the advantages and disadvantages of the algorithm. Simulation results of different reinforcement learning algorithms and parameter debugging of SAC algorithm are presented. In view of the fact that the simulation effect of reinforcement learning is better than the real test, combined with State-dependent exploration (SDE) method, parameter noise is added to improve the model training effect to make sure it can train real robots without losing performance.