Unsupervised Curricula for Visual Meta Reinforcement Learning(CARML)

The document discusses a framework for unsupervised curricula in visual meta-reinforcement learning, aimed at adapting task distributions for efficient policy training in complex environments. It emphasizes the use of a Gaussian mixture model to structure task distributions and maximize mutual information, enabling skill acquisition without predefined rewards. The proposed approach, termed CARML, showcases benefits in terms of sample efficiency during fine-tuning and evaluation of met-learning strategies on test tasks.