Abstract
In this work, we present Curled-Dreamer, a novel reinforcement learningalgorithm that integrates contrastive learning into the DreamerV3 framework toenhance performance in visual reinforcement learning tasks. By incorporatingthe contrastive loss from the CURL algorithm and a reconstruction loss fromautoencoder, Curled-Dreamer achieves significant improvements in variousDeepMind Control Suite tasks. Our extensive experiments demonstrate thatCurled-Dreamer consistently outperforms state-of-the-art algorithms, achievinghigher mean and median scores across a diverse set of tasks. The resultsindicate that the proposed approach not only accelerates learning but alsoenhances the robustness of the learned policies. This work highlights thepotential of combining different learning paradigms to achieve superiorperformance in reinforcement learning applications.