Augmented World Models Facilitate Zero-Shot Dynamics Generalization
           From a Single Offline Environment

            Philip J. Ball * 1 Cong Lu * 1 Jack Parker-Holder 1 Stephen Roberts 1

            Abstract                 problems (Dulac-Arnold et al., 2019), with the potential to
   Reinforcement learning from large-scale offline       leverage rich datasets of past experience where exploration
   datasets provides us with the ability to learn poli-    is either not feasible (e.g. a ...