Meta-World

Meta-World is an open-source benchmark of 50 distinct simulated robotic manipulation tasks performed by a single Sawyer arm in MuJoCo, built for developing and evaluating multi-task and meta-reinforcement-learning algorithms. Tasks range from reaching, pushing and pick-and-place to opening doors/drawers, pressing buttons, hammering, and assembling/disassembling pegs, and are organized into standard splits: MT1/MT10/MT25/MT50 (multi-task) and ML1/ML10/ML45 (meta-learning). Every environment follows the Gymnasium API, shares a common 4-dimensional continuous action space (end-effector Cartesian displacement plus gripper) and a fixed-length observation vector, and returns a dense shaped reward in [0, 10] together with a per-task binary success flag in the info dict. The benchmark was introduced by the Meta-World Team (Yu et al., CoRL 2019) and is now maintained by the Farama Foundation, which released the standardized 'Meta-World+' version (the current v3 environments) accepted at NeurIPS 2025 Datasets & Benchmarks. It is distributed as the MIT-licensed 'metaworld' Python package (pip install metaworld) rather than as a static downloadable dataset.