Open Source RoboticsRobots, software & learning

LIBERO

Software

Lifelong robot-learning benchmark with manipulation task suites, demonstrations and imitation-learning baselines; dataset license declarations differ across hosts. [1] [8]

Installation & first result →Official websiteGetting started

For
Researchers evaluating multitask and lifelong imitation learning [1]
Source rights
Some licenses are unresolved. Review licenses
Unknowns
operating system, compute requirements not fully established. Review gaps

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware
License
MIT · license not fully assessed for datasets · Details by part

Requirements · Research details and dates

On this page

Overview

LIBERO benchmarks knowledge transfer across simulated robot manipulation tasks. Its spatial, object and goal suites isolate different kinds of transfer, while LIBERO-100 combines 100 tasks and splits them into 90 for pretraining and 10 for downstream lifelong evaluation. This is the original benchmark repository, with its own task and training setup. [1]

Is this for you?

Known limitations
  • The README example omits imports for os and get_libero_path.
  • Dataset checks count HDF5 files rather than validate checksums or trajectories.
  • Dataset licenses conflict between the GitHub README and linked Hugging Face card.
  • Sparse reward is documented; the benchmark focuses mainly on imitation learning.

Editorial assessment from cited documentation; no runtime testing.

[1] [4] [5] [8] · Editorial assessment

What you need

Languages / runtimes
  • Python

[1]

Dependencies
  • Python 3.8.13 and PyTorch 1.11.0+cu113 in the documented environment
  • robosuite 1.4.0, robomimic 0.2.0, NumPy 1.22.4 and Gym 0.25.2
  • huggingface_hub additionally required for the Hugging Face download route

[1] [2] [4]

Installation & first result

Open the upstream getting-started guide [1]

  • Install the documented environment and select a task suite.
  • For demonstrations, install huggingface_hub and request one suite with --datasets and --use-huggingface.
  • Create the environment, set its seed and benchmark initial state, then inspect a short rollout before training.

Editorial assessment from cited documentation; no runtime testing.

[1] [3] [4] · Editorial assessment

Installation

Use the pinned README environment and requirements before pip install -e .; modern dependency compatibility is unverified.

[1]

Hardware, models & data

Inputs and outputs

Task descriptions, BDDL scenes, fixed initial states and demonstrations support policy training and success evaluation.

[1]

Project notes

Start with one suite and the documented environment

The README uses Python 3.8.13 and PyTorch 1.11.0 with CUDA 11.3 packages. Its requirements pin robosuite 1.4.0, robomimic 0.2.0, NumPy 1.22.4 and Gym 0.25.2. Follow that environment when evaluating this revision; compatibility with current releases of those packages is not established here. [1] [2]

For a limited data trial, request one suite, for example python benchmark_scripts/download_libero_datasets.py --datasets libero_spatial --use-huggingface. The downloader requires huggingface_hub for this route, but it is absent from the pinned requirements list, so install it separately. Without a suite selection, the script requests all datasets. Its original-link route warns that those links may expire and offers the Hugging Face alternative. [2] [3] [4]

Preserve the benchmark initial state

The README's task example creates an off-screen environment, sets a seed, resets it and applies a supplied benchmark initial state before stepping. Keep that state selection when comparing evaluations. The snippet also needs import os and from libero.libero import get_libero_path, which it uses without importing; those imports are a source-based correction, not a tested reproduction. [1] [5]

A ten-step zero-action rollout only exercises environment stepping. It does not demonstrate task success or policy quality. The README documents sparse success rewards and mainly focuses on lifelong imitation learning. Choose a policy and learning algorithm only after confirming the task scene and demonstrations load as intended. [1]

The downloader's completeness check counts HDF5 files—ten for most suites and ninety for LIBERO-90. It does not verify file checksums or demonstration contents, and it checks all suites even when you intentionally downloaded one. Missing-suite messages therefore need to be interpreted against your selected download. [3] [4]

Dataset permission is unresolved across hosts

The GitHub README assigns CC-BY-4.0 to datasets, while the linked Hugging Face dataset card declares Apache-2.0. We have retained both declarations and leave dataset license status unresolved; do not infer a permission choice from which downloader you use. The repository code and documentation have MIT terms. [1] [7] [8]

This profile reviews commit 8f1084e3132a39270c3a13ebe37270a43ece2a01, with no GitHub release listed at review. We inspected a scoped source set and the public dataset listing/card, without downloading demonstrations, installing the simulator, training a policy or measuring benchmark performance. [6] [8]

What we checked

Source researched only; no installation, training, benchmark or physical hardware test performed. No affiliate links or reused media.

Not yet established

  • Requirements: operating system, host computer
  • Version-specific hardware compatibility (ecosystem tags alone do not establish it)

Updates

Upstream last updated
Mar 15, 2025 [6]
Latest release
Not established
Profile updated
Oct 17, 2026
Sources last checked
Oct 17, 2026
Next review
Nov 17, 2026
Reviewed revision
commit snapshot; no GitHub releases listed / 8f1084e3132a39270c3a13ebe37270a43ece2a01 · software repository

Source files and licenses

MIT benchmark source; dataset rights unresolved due conflicting upstream declarations.

LayerAvailability / license
softwareavailable · MIT
Repository materials only; external components retain their own terms. [7]
datasetsavailable · License not fully assessed
GitHub README: CC-BY-4.0; linked HF card: Apache-2.0. Conflict unresolved. Public listing inspected; dataset payloads not downloaded. [1] [8]
documentationavailable · MIT
Repository materials only; external components retain their own terms. [7]

The scoped benchmark workflow trains/evaluates user-selected policies rather than reviewing a pretrained release.

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.7.0-draft.5

Change history