LIBERO

Lifelong robot learning benchmark with 130 procedural tasks

License

MIT

Overview

A benchmark for lifelong robot learning featuring 130 procedurally-generated tasks across manipulation, spatial, object, and goal domains. Built on MuJoCo physics simulation with standardized evaluation protocols for knowledge transfer and continual learning research.

LIBERO is a comprehensive benchmark suite designed to evaluate and advance lifelong robot learning capabilities through a diverse collection of 130 procedurally-generated manipulation tasks. The benchmark addresses the critical challenge of enabling robots to continuously acquire new skills while retaining previously learned behaviors, organizing tasks into four task suites: LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-100. Built on the MuJoCo physics engine, LIBERO provides a standardized simulation environment with consistent evaluation protocols, enabling reproducible comparisons across different learning algorithms. The procedural generation system ensures task diversity while maintaining controllable difficulty progression, allowing researchers to systematically study how robots can accumulate knowledge over extended learning periods.

Why it matters

LIBERO is the first comprehensive benchmark specifically designed for lifelong robot learning. With 130 procedurally-generated tasks across four domains, it provides the standardized evaluation needed to advance continual and lifelong learning for manipulation.

Problem it solves

Robots deployed in the real world need to continuously learn new tasks without forgetting previous skills. LIBERO provides a standardized benchmark to measure whether learning algorithms can actually achieve lifelong learning for manipulation.

Key takeaways

• 130 procedurally-generated manipulation tasks across 4 domains
• Specifically designed to evaluate lifelong and continual learning
• Domains cover spatial, object, goal, and long-horizon manipulation
• Built on MuJoCo with standardized evaluation protocols
• Tests whether agents can learn new tasks without forgetting old ones

Who it is for

• Researchers in continual and lifelong learning for robotics
• Teams building adaptive manipulation policies
• Developers evaluating catastrophic forgetting in robot policies
• Students studying transfer learning in manipulation

Limitations

• Simulated environment only — no real-robot transfer evaluation
• MuJoCo-based tasks may not reflect real-world complexity
• Procedural generation can produce similar-feeling tasks
• Focused on manipulation — no locomotion or navigation tasks