Most deep reinforcement learning algorithms (used in robotics, video games, and driving simulators among other settings) use a reward function to learn visual features. Visual features and a control policy are learned jointly. However, recent work suggests that visual features for RL can be learned without rewards. Stooke et al. propose a new unsupervised learning task, Augmented Temporal Contrast (ATC), which makes a model link observations from nearby timesteps within the same trajectory. By training an encoder exclusively using ATC, they demonstrate that reinforcement learning can be decoupled from reinforcement learning to match to improve upon the performance of end-to-end RL systems.