LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning

Liu, Zhu, Gao, Feng, Liu, Zhu, Stone · 2023 · arXiv preprint · arXiv:2306.03310 · PDF

Dongyu supplement. Added as a candidate missing/adjacent dataset paper after searching primary arXiv sources. This is a draft summary for triage, not a full paper read.

One-liner. A language-conditioned lifelong robot learning benchmark with four task suites, 130 tasks, and human teleoperated demonstrations.

Setup

• Datasets / benchmarks: Introduces LIBERO: four benchmark task suites with 130 tasks total, high-quality human-teleoperated demonstration data for all tasks, and public code/datasets.
• Hardware / simulator: Simulation benchmark with procedural generation pipeline; task suites are designed for knowledge transfer and sample-efficient learning.

Method

Benchmark for lifelong decision-making and transfer of declarative/procedural knowledge in robot manipulation.

Why it matters for the map

Important for language-conditioned manipulation benchmarks and task-suite organization; useful comparator for any new benchmark proposal.

Limitations / open questions

Simulation/vision/language benchmark; not centered on tactile/audio/force/thermal sensing.

Source note

arXiv lines 38-41 report benchmark goals, four suites, 130 tasks, demos, and public code/datasets.