Скачать или смотреть Reinforcement Learning for Control course - Lecture 18

In this lecture, we build a complete MATLAB and Simulink workflow for training a DDPG agent to control UAV altitude using the Reinforcement Learning Toolbox. The session is fully hands-on and demonstrates each step in detail, starting from the dynamic model and ending with evaluation of the trained controller.

We begin by creating a simple vertical UAV model in Simulink that captures mass, gravity, thrust generation and actuator limits. We then define the observation space using altitude and vertical velocity. The continuous action is mapped to a normalized throttle command that is scaled and limited before reaching the plant.

Next, we design the reward function for stable altitude tracking. The RL environment is assembled using rlSimulinkEnv, with appropriate sample times, episode durations, and a reset function that introduces randomized initial conditions for more robust learning.

We then construct the DDPG agent by creating actor and critic networks with fully connected layers and ReLU activations. Target networks and experience replay buffers are configured to stabilize learning, and Ornstein–Uhlenbeck noise is added to enable smooth exploration in continuous action spaces. Training is performed using carefully chosen options for episodes, steps, stopping criteria, and checkpointing.

Throughout the lecture, we examine learning curves, episode rewards and trajectory tracking to understand the training progress. After training, we run deterministic simulations to verify closed-loop performance, test step changes in altitude commands. The lecture concludes with instructions for saving the trained agent and exporting the policy.

If you have any questions or feedback, feel free to reach out through my website: https://sites.google.com/view/mbmohiu...