Reinforcement Learning for Real-World Robotics

The intersection of expressive, general-purpose function approximators, such as neural networks, with general-purpose model-free reinforcement learning holds the promise of automating the acquisition of a wide range of robotic skills: reinforcement learning provides the formalism for reasoning about sequential decision making, while large neural networks can provide a general representation for any behavior. However, learning behaviors from scratch via deep reinforcement learning in real-world robotic domains has proven to be challenging, and only recently, we have started to see progress mainly due to improved sample complexity and stability of such methods. In this talk, I will first describe my work at the intersection of robotics and deep reinforcement learning, focusing on learning robot skills from scratch in the absence of prior knowledge, and then discuss the biggest limitations and most promising future directions.

Dr. Tuomas Haarnoja is a research scientist at Google Deepmind. He did his doctoral thesis with Pieter Abbeel and Sergey Levine at the Berkeley AI Research Lab (BAIR), where he became well known for his work on maximum entropy reinforcement learning, which provides a theoretically grounded framework for learning stochastic policies that are both sample efficient and reliable, and its applications to robotic manipulation and locomotion.