The D4RL benchmark is introduced to evaluate progress in offline reinforcement learning (RL), where policies are learned from static datasets. This benchmark is designed to reflect real-world applications of offline RL, focusing on dataset collection and properties such as narrow data distributions, multitask data, and human demonstrations. The benchmark includes a variety of tasks and datasets, such as Maze2D, AntMaze, Gym-MuJoCo, Adroit, FrankaKitchen, Flow, and Offline CARLA, each designed to challenge existing RL algorithms. These tasks are based on simulated environments that have been widely used in the research community, allowing for reproducible and accessible evaluations. The benchmark provides a comprehensive evaluation of existing algorithms, including both online and offline RL methods, and includes a standardized protocol for comparison. The results show that many algorithms struggle with tasks that have properties relevant to real-world applications, such as sparse rewards and non-representable policies. The benchmark also highlights the importance of realistic data collection procedures in evaluating RL algorithms. The D4RL benchmark serves as a common starting point for the community to identify shortcomings in existing offline RL methods and to advance research in this emerging area.The D4RL benchmark is introduced to evaluate progress in offline reinforcement learning (RL), where policies are learned from static datasets. This benchmark is designed to reflect real-world applications of offline RL, focusing on dataset collection and properties such as narrow data distributions, multitask data, and human demonstrations. The benchmark includes a variety of tasks and datasets, such as Maze2D, AntMaze, Gym-MuJoCo, Adroit, FrankaKitchen, Flow, and Offline CARLA, each designed to challenge existing RL algorithms. These tasks are based on simulated environments that have been widely used in the research community, allowing for reproducible and accessible evaluations. The benchmark provides a comprehensive evaluation of existing algorithms, including both online and offline RL methods, and includes a standardized protocol for comparison. The results show that many algorithms struggle with tasks that have properties relevant to real-world applications, such as sparse rewards and non-representable policies. The benchmark also highlights the importance of realistic data collection procedures in evaluating RL algorithms. The D4RL benchmark serves as a common starting point for the community to identify shortcomings in existing offline RL methods and to advance research in this emerging area.