Title: Deep Learning for Robots and Agents
Deep learning has the potential to revolutionise robotics through end-to-end learning, using the same methods that have enabled human-level speech and image recognition. But robot domains hold special challenges because of their interactive, sequential nature. Recent research from DeepMind has begun to answer some of these challenges through novel architectures and algorithms for deep reinforcement learning. I will describe a general approach for continual and transfer learning, and demonstrate a method for navigation in complex environments.