LIBERO is a lifelong robot learning benchmark for studying knowledge transfer.

In the future, robots will be pretrained with a diverse set of skills in factory before deployment. However, we believe it is also a must for them to learn with their human users throughout their lifetime to really become a personalized embodied agent. This pertains to the paradigm of lifelong learning in decision making (LLDM). While LLDM is an important research question, currently the community lacks good testbeds for studying LLDM. To this end, we propose LIBERO.

The full release is coming soon!

Overview

Procedural generation pipeline from human activity datasets

130 tasks

65,000 high-quality demonstrations for sample-efficient leanring

Build on top of a modularized robot simulator, Robosuite

Multiple implentation of neural architectures, lifelong learning algorithms.

The LIBERO Team

Bo Liu

Yifeng Zhu

Chongkai Gao

Yihao Feng

Qiang Liu

Yuke Zhu

Peter Stone