D4RL

D4RL (Datasets for Deep Data-Driven Reinforcement Learning) is the most widely cited benchmark suite for offline reinforcement learning in robotics. Introduced in 2020 by researchers from Google, UC Berkeley, and Stanford, D4RL provides standardized datasets and evaluation protocols for offline RL algorithms across diverse simulated robotic domains. The benchmark covers multiple environments: MuJoCo locomotion (HalfCheetah, Hopper, Walker2D), the Adroit dexterous manipulation hand (door, hammer, pen, relocate), FrankaKitchen manipulation, CARLA autonomous driving, Flow traffic control, and Maze2D navigation. Each domain includes multiple dataset types: hand-designed controllers, human demonstrations, and multi-task data. D4RL revealed critical deficiencies in offline RL algorithms and drove the development of CQL, IQL, and other major offline RL methods. It has been cited in thousands of papers and remains the standard evaluation benchmark for offline RL research. D4RL is now maintained as part of the Farama Foundation's ecosystem, alongside Gymnasium and other core RL libraries.