LIBERO is a benchmark for lifelong learning in robot manipulation, where a robot learns tasks one after another and should carry knowledge forward. The authors argue that lifelong decision-making must transfer procedural knowledge (how to act) as well as declarative knowledge (what things are). A procedural generator can in principle make unlimited tasks, and every task comes with human teleoperated demonstrations.
Findings
- Four task suites, 130 tasks in total, organised around five questions: which kind of knowledge transfers, policy architecture, lifelong learning algorithms, robustness to task order, and the effect of pretraining.
- Plain sequential fine-tuning beats existing lifelong learning methods on forward transfer.
- No single visual encoder is best for every kind of transfer.
- Naive supervised pretraining can hurt later lifelong learning.
Limits
- Abstract-level reading only.
Relevance
One of the few benchmarks that measures learning over a sequence. Its “lifelong” means a run of new tasks; a robot left alone for weeks also faces a world that changes under the same task, a gap that long-term autonomy work could fill. Used by Agentic Robot and bundled in LeRobot; see embodied-ai-benchmarks.
Links
- arXiv
- Subfield: Robot learning