Abstract

Join us on Zoom

The goal of offline reinforcement learning is to learn a policy from a fixed dataset, without further interactions with the environment. This setting will be an increasingly more important paradigm for real-world applications of reinforcement learning such as robotics, in which data collection is slow and potentially dangerous. In this talk, we will discuss the challenges of applying existing off-policy algorithms on static datasets and the reasonings behind the objectives of offline RL. We will then introduce our approach Policy in the Latent Action Space (PLAS) which naturally satisfies the objectives. Our method is evaluated on continuous control benchmarks in simulation and the cloth-sliding task with a physical robot. We demonstrate that our method provides competitive performance consistently across various continuous control tasks and different types of datasets, outperforming previous offline reinforcement learning methods with explicit constraints.

Bio: Wenxuan Zhou is a Ph.D. student at the Robotics Institute at Carnegie Mellon University, advised by Prof. David Held. Her research interests lie at the intersection of robotics and reinforcement learning. Previously, she received her Master’s degree in Robotics at CMU advised by Prof. Abhinav Gupta. Prior to that, she obtained her dual B.S. degrees in Electrical and Computer Engineering from Shanghai Jiao Tong University and Mechanical Engineering from the University of Michigan. She will be joining DeepMind as an intern in Summer 2021.