Back to Benchmarks
LIBERO
LIBERO benchmark for lifelong robot learning: 130 manipulation tasks on RoboSuite (MuJoCo), 65K expert demos, and results such as InternVLA-M1 at 95.9%.
Lifelong roBotic lEarning benchmaRk. 130 tasks, spatial/object/goal suites.
Overview
LIBERO evaluates lifelong learning: agents that must learn new tasks sequentially without forgetting. Built on RoboSuite (MuJoCo). Three suites: spatial, object, goal. 65K expert demos available.
Key Results
- InternVLA-M1: 95.9% success
Official Links
- libero-project.github.io — Project site
- LIBERO Dataset — Our datasets page
Benchmarks compare policies in sim; our eval loop scores them on physical rigs.