LIBERO: Lifelong Robot Learning Benchmark

A curated guide to LIBERO — the 130-task, 65K-demo benchmark that has become the canonical lifelong imitation learning evaluation target for robot manipulation policies.

TL;DR

Metric	Value
Task count	130 tasks across 4 suites (Spatial, Object, Goal, LIBERO-100)
Robots	Franka Emika Panda (simulated via RoboSuite)
Modalities	RGB (2 cameras), proprioception, gripper, language
License	MIT
Size	~65,000 demonstrations, ~40 GB HDF5
Simulator	RoboSuite (MuJoCo)

What is LIBERO?

LIBERO (Lifelong Robot Learning Benchmark) is a research benchmark introduced by the UT Austin Robot Perception and Learning Lab to study how manipulation policies transfer knowledge across a continual stream of tasks. Rather than asking a single policy to memorize a fixed set of demonstrations, LIBERO organizes 130 tabletop manipulation tasks into four lifelong learning suites that each isolate a different transfer axis: LIBERO-Spatial varies object placement, LIBERO-Object varies object identity, LIBERO-Goal varies the task goal, and LIBERO-100 (split into LIBERO-90 pretraining plus LIBERO-10 evaluation) tests long-horizon compositions of everyday skills.

Every task ships with 50 human teleoperated demonstrations collected in the RoboSuite simulator on a Franka Emika Panda, producing roughly 65,000 trajectories in total. Observations include two RGB camera streams (agent-view and in-hand), proprioceptive joint state, gripper state, and a natural language instruction string. Ground-truth object and scene metadata are also exposed so researchers can study privileged-information baselines, procedural generalization, or sim-to-real distribution shift.

Because the benchmark is drop-in compatible with RoboSuite and ships a Gym-style evaluation harness, LIBERO has become the de-facto lifelong learning evaluation target for vision-language-action models. OpenVLA, Octo, Diffusion Policy, ACT, and BC-Z all publish LIBERO numbers, which makes it the fastest way to place a new policy on a shared leaderboard.

How to download & load

The canonical distribution is HDF5, and there is also a modified RLDS fork used by the OpenVLA team. Install and pull the datasets with:

# Clone the benchmark code
git clone https://github.com/Lifelong-Robot-Learning/LIBERO.git
cd LIBERO && pip install -e .

# Download the HDF5 datasets (~40 GB)
python benchmark_scripts/download_libero_datasets.py --datasets libero_spatial libero_object libero_goal libero_100

# Or pull the RLDS fork used by OpenVLA / Octo
pip install datasets
python -c "from datasets import load_dataset; d = load_dataset('openvla/modified_libero_rlds', split='train'); print(d[0].keys())"

For fine-tuning, the RLDS variant plugs directly into the OpenVLA and Octo training scripts. For evaluation, use the shipped libero/libero/benchmark/<suite>.py entry points — they load the scene, replay the language instruction, and score success with the original task-specific reward.

Common use cases & pairings

• Fine-tuning VLAs. OpenVLA and Octo both publish LIBERO-tuned checkpoints — the suite is the fastest way to validate that a VLA fine-tune is not catastrophically forgetting.
• Diffusion Policy and ACT baselines. The 50-demo-per-task setting is well-matched to diffusion and transformer policies; the leaderboard makes it easy to compare against published numbers.
• Continual learning research. The four suites are the only publicly scaled robot manipulation benchmark where you can measure forgetting, forward transfer, and backward transfer with large sample sizes.
• Procedural augmentation. Because scenes are defined in MuJoCo XML, researchers frequently pair LIBERO with MimicGen or domain randomization to study data scaling laws.

Benchmarks & leaderboards

Public leaderboards track success rate per suite (LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, LIBERO-10) averaged over 500 evaluation rollouts. Current state of the art sits in the 95%+ range on the first three suites and 50-60% on LIBERO-10. See the Papers with Code LIBERO leaderboard and the project page for canonical numbers.