Advancing Robust Controllers with Reinforcement Learning | ETH Zürich Real World Robotics Tutorial 6

In this video, Professor Robert Katzschmann recaps neural networks and their application in image classification. The discussion then delves into reinforcement learning, exploring agent-environment interactions, rewards, actions, and observations. Concepts like policy, value function, and Q function are introduced, leading to a demonstration of the Q learning algorithm. The video also covers the actor-critic model and state-of-the-art algorithms using this approach. The importance of learning from experts is emphasized, with examples from human input and additional information. The simplest imitation learning approach is presented, along with its limitations. The video concludes by addressing the challenges of training in simulation versus real-world dynamics and strategies to mitigate them.