LIBERO

LIBERO is a benchmark for studying knowledge transfer in multitask and lifelong robot manipulation, released with high-quality human-teleoperated demonstration data for all of its tasks. It contains 130 tasks organized into four suites: LIBERO-Spatial (10 tasks), LIBERO-Object (10 tasks), LIBERO-Goal (10 tasks), and LIBERO-100 (100 tasks, split into LIBERO-90 for pretraining and LIBERO-10/LIBERO-Long for downstream lifelong learning). The authors provide 50 human-teleoperated demonstrations for every task, giving 6,500 demonstrations total, each a variable-length trajectory of a simulated 7-DoF manipulator collected in the robosuite/MuJoCo simulator. Every task is defined by a BDDL file that specifies a natural-language instruction, an initial scene layout, and a symbolic goal predicate; the simulator emits a sparse +1 reward only when that goal predicate is satisfied. Demonstrations are stored per task as robomimic-format HDF5 files with two 128x128 RGB camera streams (third-person agentview and wrist eye-in-hand), proprioceptive robot state, and 7-dimensional continuous actions. The codebase is MIT-licensed and the demonstration datasets are released under CC BY 4.0.