LIBERO

LIBERO (Lifelong Robot Learning Benchmark) is a benchmark suite for studying knowledge transfer in lifelong robot manipulation, developed by researchers at UT Austin. Provides a standardized framework for evaluating how robot policies retain and transfer knowledge across sequentially learned tasks. The benchmark comprises 130 manipulation tasks across five suites: LIBERO-Spatial, LIBERO-Object, LIBERO-Goal (10 each, isolating specific distribution shifts), LIBERO-90 (pretraining), and LIBERO-10 (evaluation). Data includes RGB images, proprioception, language specifications, and PDDL scene descriptions, totaling ~100 GB. Published at NeurIPS 2023 (Dataset and Benchmark Track). Licensed under MIT (code) and CC BY 4.0 (datasets). A critical evaluation tool for generalist robot policy research.