Task-Aware Virtual Training: Enhancing Generalization in Meta-Reinforcement Learning for Out-of-Distribution Tasks

Learn task representations and train on virtual tasks to generalize beyond the training distribution.

TAVT learns task representations using a task-distance metric and generates virtual experience that preserves task characteristics. State regularization addresses overestimation in environments with varying states; evaluations cover MuJoCo and MetaWorld tasks outside the training distribution.