LIBERO

Lifelong robot-manipulation benchmark (NeurIPS 2023 D&B): 130 language-conditioned tasks across four suites probing knowledge transfer in declarative and procedural skills. Pre-window seed that became the default VLA manipulation benchmark — OpenVLA, π0, and most 2025 vision-language-action papers report LIBERO numbers. From Yuke Zhu's Robot Perception & Learning lab.