Open Source RoboticsRobots, software & learning

LIBERO-PRO

Software

LIBERO evaluation extension that perturbs objects, positions, language, task goals and environments to study policy generalization in simulation. [1]

Installation & first result →Official websiteGetting started

For
Researchers integrating a vision-language-action policy with a simulation evaluator [1]
Source rights
Named licenses are documented by part. 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

Requirements · Research details and dates

On this page

Overview

LIBERO-PRO extends LIBERO manipulation evaluations with changes to objects, positions, language, task goals and environments. It is for researchers asking whether a policy still succeeds when a familiar simulated task changes. Its role is evaluation infrastructure; a policy checkpoint and its execution code are separate inputs. [1]

Is this for you?

Known limitations
  • Task perturbation cannot be combined with other perturbations.
  • Missing perturbation configuration paths can be skipped while returning unchanged task content.
  • README evaluator code is an integration sketch, not a verified standalone command.
  • ASPIRE uses a separate fork; compatibility with this upstream snapshot is unverified.

Editorial assessment from cited documentation; no runtime testing.

[1] [4] [8] · Editorial assessment

What you need

Languages / runtimes
  • Python

[4] [5]

Dependencies
  • Python 3.8.13 in the documented environment
  • PyTorch 1.11.0+cu113
  • robosuite 1.4.0; robomimic 0.2.0; bddl 1.0.1
  • Separately downloaded BDDL and initial-state files plus a policy evaluator/checkpoint

[1] [2]

Host computer

The documented environment uses CUDA 11.3 PyTorch wheels; policy checkpoint memory, rendering and evaluation costs are not fully specified.

[1]

Installation & first result

Open the upstream getting-started guide [1] · Project documentation [1]

  • Create the documented Python environment and install requirements, PyTorch and the editable package.
  • Download matching BDDL and initial-state folders from the linked dataset.
  • Check paths and enable one perturbation for a first experiment.
  • Adapt the external policy evaluator and inspect generated task files before reporting results.

[1] [3] [4]

Installation

Follow the README Python 3.8.13 environment and explicit dependency pins; policy setup is separate.

[1] [2]

Hardware, models & data

Inputs and outputs

BDDL tasks, initial states, YAML perturbation configuration and a policy evaluator produce modified evaluation tasks and policy results.

[1] [3] [4]

File formats
  • BDDL
  • YAML
  • init files

[1] [3]

Evaluation

Record original versus perturbed task results with the policy, seed, perturbation configuration and matching initial-state files.

Editorial assessment from cited documentation; no runtime testing.

[1] [4] · Editorial assessment

Project notes

Keep the upstream and ASPIRE fork distinct

This profile reviews the original Zxy-MLlab/LIBERO-PRO repository at a commit snapshot, with no GitHub release listed at review. ASPIRE led us to the uynitsuj/LIBERO-PRO fork, whose README directs readers to this original project. We retain that fork as a separate discovery relationship and have not verified that the two revisions are interchangeable. [6] [8]

Reconstruct the documented environment carefully

The README specifies Python 3.8.13 and PyTorch 1.11.0 with CUDA 11.3 wheels. Its requirements pin robosuite 1.4.0, robomimic 0.2.0 and bddl 1.0.1. The quick start also requires BDDL task files and initial-state files from the linked Hugging Face dataset, placed into matching directories. Installing the package alone does not supply that complete evaluation setup. [1] [2] [5]

The OpenVLA section asks readers to modify an external evaluator. Its example contains from LIBERO-PRO import perturbation, which is not valid Python import syntax because the module name contains a hyphen. Treat this section as an integration sketch requiring correction, not as a command sequence we have reproduced. The inspected sources do not establish a complete GPU-memory requirement or evaluation cost for a chosen policy. [1]

Verify that the perturbation actually happened

Begin with one perturbation and inspect its generated task files before comparing policy scores. The README describes both a single-perturbation mode and optional combinations; the implementation explicitly rejects combining task perturbation with any other type. The checked-in YAML enables environment perturbation, so do not assume it matches the README's task-perturbation example. [1] [3] [4]

Several implementation branches print a message and skip a perturbation if its configuration path is missing. Other branches return the original task when no replacement is configured. A completed run therefore does not by itself show that the intended task changed. Preserve the configuration, seed, generated BDDL and matching initial-state files alongside results. [4]

Upstream also reports unintended object movement after some environment replacements and says main_table worked in its tests. The inspected implementation instead fixes the replacement target to living_room_table. We have not resolved the runtime implications of that mismatch. Check initialized object positions before attributing score changes solely to the intended environment change. [1] [4]

The code license is MIT; the README separately declares CC-BY-4.0 for datasets. We preserved selected pinned documentation, configuration and implementation files, not the complete asset collection or external dataset. We have not installed the environment, run a policy or reproduced the published benchmark scores. [7] [1]

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
  • Version-specific hardware compatibility (ecosystem tags alone do not establish it)

Updates

Upstream last updated
Mar 23, 2026 [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 / eafdb809426b13153aa1e4c42d6601844217dfec · software repository

Source files and licenses

Useful source-researched evaluation extension with explicit integration gaps; fork identity and external data remain separately scoped.

LayerAvailability / license
softwareavailable · MIT
Repository materials only; external components retain their own terms. [7]
datasetsavailable · CC-BY-4.0 (declared by upstream README)
Dataset license declaration inspected; external payloads not downloaded or independently audited. [1]
documentationavailable · MIT
Repository materials only; external components retain their own terms. [7]

Policy checkpoints are separate evaluator inputs; this profile does not audit their licenses.

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