Reinforcement learning in robotics: A survey

This survey connects the reinforcement learning and robotics research communities by reviewing work on using RL to generate robot behaviors. It frames robotics as a source of inspiration, impact, and validation for RL, and organizes the literature around the challenges of applying RL to real robots. The authors emphasize the roles of algorithms, representations, and prior knowledge in taming the complexity of the domain, with a particular focus on the design choices between model-based and model-free methods and between value-function-based and policy-search methods.

By analyzing a simple problem in detail, the paper demonstrates how RL approaches can be applied profitably in robotics and highlights notable successes along the way. It matters because it consolidates scattered advances into a coherent picture of what works and why, identifies open questions, and points to the substantial future potential of strengthening the links between reinforcement learning and robotics.