CURL: Contrastive Unsupervised Representations for Reinforcement\n Learning

Abstract

We present CURL: Contrastive Unsupervised Representations for Reinforcement\nLearning. CURL extracts high-level features from raw pixels using contrastive\nlearning and performs off-policy control on top of the extracted features. CURL\noutperforms prior pixel-based methods, both model-based and model-free, on\ncomplex tasks in the DeepMind Control Suite and Atari Games showing 1.9x and\n1.2x performance gains at the 100K environment and interaction steps benchmarks\nrespectively. On the DeepMind Control Suite, CURL is the first image-based\nalgorithm to nearly match the sample-efficiency of methods that use state-based\nfeatures. Our code is open-sourced and available at\nhttps://github.com/MishaLaskin/curl.\n