LIBERO — Lifelong Learning for Robot Manipulation
LIBERO is a lifelong robot learning benchmark and simulation dataset developed at UT Austin covering four task suites designed to evaluate continual learning in robot manipulation. Released in 2023 under MIT license, it contains 130,000 demonstration episodes across LIBERO-Spatial (same objects, different locations), LIBERO-Object (same locations, different objects), LIBERO-Goal (same scene, different goal states), and LIBERO-Long (long-horizon sequential tasks). LIBERO addresses a fundamental challenge in robot learning: policies typically forget previous tasks when trained on new ones (catastrophic forgetting). The four benchmark suites systematically isolate different sources of task variation to study how well lifelong learning algorithms retain previous skills. LIBERO is distributed in HDF5 format and available via the LeRobot library. It is the standard benchmark for continual and lifelong robot learning research.
| Year | 2023 |
|---|---|
| Episodes | 130,000 |
| Embodiments | Franka Panda (simulated) |
| Modalities | rgb, proprioception |
| Task categories | manipulation, pick-and-place, long-horizon |
| Data format | hdf5, lerobot |
| License | MIT |
| Access | open — commercial use permitted |
| Maintainer | UT Austin, Collaborative AI & Robotics Lab |
| Origin country | US |
What is it?
LIBERO is a lifelong robot learning benchmark developed at UT Austin covering four task suites designed to evaluate continual learning in robot manipulation. Released in 2023 under MIT license, it contains 130,000 demonstration episodes across LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-Long. LIBERO addresses catastrophic forgetting — the tendency of neural networks to forget previous tasks when trained on new ones.
Who is it for?
Researchers working on continual learning, lifelong learning, and multi-task robot manipulation. Particularly valuable for teams developing algorithms that accumulate skills over time rather than being retrained from scratch for each new task — a critical requirement for real-world robot deployment.
Key specifications
- Episodes: 130,000 demonstrations
- Task suites: LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, LIBERO-Long
- Robot platform: Simulated Franka Panda
- Format: HDF5, LeRobot
- License: MIT — commercial use permitted
- Access: Open — Hugging Face and GitHub
How it compares
The standard benchmark for continual robot learning. Meta-World covers more task types (50) but focuses on multi-task RL. RoboCasa covers more household tasks but doesn't isolate sources of task variation for continual learning study. LIBERO's four-suite structure is its defining contribution.
Limitations and access notes
Simulation-only with Franka Panda. The four suites are designed for controlled experimental comparison rather than real-world diversity. MIT license permits unrestricted commercial use.
Linked professions
Frequently asked questions
What is catastrophic forgetting and why does LIBERO address it?
Catastrophic forgetting is when neural networks lose performance on previously learned tasks when trained on new ones. LIBERO's four structured task suites isolate different sources of task variation, enabling study of whether continual learning algorithms successfully retain prior skills while acquiring new ones.
What are the four LIBERO task suites?
LIBERO-Spatial tests generalisation to different object locations. LIBERO-Object tests generalisation to different objects in the same locations. LIBERO-Goal tests generalisation to different goal states in the same scene. LIBERO-Long tests long-horizon sequential task completion requiring multiple chained manipulation steps.
Can LIBERO be used commercially?
Yes. LIBERO is MIT licensed, permitting unrestricted commercial use.
How do I access LIBERO?
LIBERO is available on Hugging Face (lerobot/libero_goal and related datasets) and via GitHub at github.com/Lifelong-Robot-Learning/LIBERO. No registration is required.
What algorithms have been benchmarked on LIBERO?
LIBERO has benchmarked EWC, PackNet, Progressive Neural Networks, Experience Replay, LOTUS, and various multi-task policy architectures. It is the reference for continual robot learning papers.