Robotic World Model: Toward Real-World Online Reinforcement Learning

Abstract

In this talk, I will present our recent work on developing a neural network simulator, Robotic World Model (RWM), for robust policy optimization in robotics. RWM is designed to be a fast and accurate simulator that can be used for online reinforcement learning in the real world. I will discuss the architecture of RWM, its training process, and its performance in various robotic tasks. I will also extend the discussion to include uncertainty awareness in RWM model training, which can be leveraged for policy regularization when optimized in imagination.