Robot Learning 2026 – Lecture 4: Reinforcement Learning I | ETH Zürich

Reinforcement learning enables agents to optimize policies through trial-and-error, overcoming the performance ceilings inherent in imitation learning. Central to this approach is the Markov decision process, where value functions and the Bellman equation compress infinite-horizon planning into efficient, one-step recursive updates. While tabular methods provide exact solutions, deep reinforcement learning—exemplified by DQN—leverages neural networks and experience replay to handle complex, high-dimensional inputs like raw pixels. For robotics, where continuous action spaces pose significant optimization challenges, methods such as cross-entropy sampling and actor-critic architectures like DDPG allow agents to learn optimal behaviors without requiring explicit dynamics models. These advancements facilitate autonomous skill acquisition, such as robots learning to rearrange cluttered objects before grasping, by maximizing reward signals rather than relying on hand-coded heuristics.