LIBERO Benchmark

A simulated benchmark of 130 tabletop manipulation tasks, one of the most commonly reported in VLA papers.

LIBERO was proposed by Bo Liu, Yuke Zhu, Peter Stone, and colleagues, published at the NeurIPS 2023 Datasets and Benchmarks track, and originally designed to study knowledge transfer in lifelong learning. It's built on robosuite (which runs on MuJoCo underneath), and procedurally generates 130 tasks with language instructions, split into four suites: Spatial (varying object placement), Object (varying which object), and Goal (varying the goal), each with 10 tasks; LIBERO-100 splits further into LIBERO-90, used for pretraining, and 10 long-horizon tasks called LIBERO-Long (also known as LIBERO-10). Each task comes with 50 human-teleoperated demonstrations. It later became the most commonly used leaderboard for VLA fine-tuning, and leading methods now exceed 97% average success rate, which has led to harder variants with added perturbations, LIBERO-Plus and LIBERO-PRO.

ExampleOpenVLA-OFT reaches 97.1% average success rate across LIBERO's four suites, compared with 76.5% for the original OpenVLA.

Also called

Benchmarking Knowledge Transfer for Lifelong Robot Learning, LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, LIBERO-Long (LIBERO-10), LIBERO-90, LIBERO-100

Related

Benchmark · Benchmark Saturation · LIBERO-Plus · LIBERO-PRO · robosuite · OpenVLA-OFT

Sources

LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning (arXiv 2306.03310)
LIBERO GitHub repository
OpenVLA-OFT (arXiv 2502.19645)

As of

2026-09