zhouxueyang/LIBERO-Pro2.0-Eval

LIBERO-Pro2.0: A Benchmark for Evaluating the Robustness of Embodied Foundation Models

🏗️ Repo | 🌐 Website | 🤗 Assets | 📁 Training Dataset

🔥 Overview

This repository provides the evaluation resources for LIBERO-Pro2.0, a robustness benchmark for embodied foundation models. It evaluates generalization under static distribution shifts and adaptation to interventions during task execution, covering 22 static and 20 runtime perturbation types across six domains.

The release includes BDDL tasks, initial states, simulator assets, evaluation service/client code and resumable batch runners. GitHub maintains the source; LIBERO-Pro2.0-Train provides HDF5 demonstrations, training metadata and trajectory source indices. Training trajectories are not required to evaluate a pretrained policy.

📊 LIBERO-Pro2.0 Benchmark

6 Perturbation Domains

The released catalogue contains 1,688 task instances across LIBERO-Spatial, LIBERO-Object, LIBERO-Goal and LIBERO-10: 40 BASE, 863 static and 785 runtime instances.

1. Condition - Object poses, task geometry, receiving regions and interaction constraints.
2. Environment - Object instances, distractors, clutter, appearance and scene context.
3. Observation - Camera settings, sensor quality, frame streams and modality alignment.
4. Execution - Control timing, command delivery, action chunks and temporary interventions.
5. Robot - Initialization, calibration, joint state, grasp failures and recoverable drops.
6. Language - Instruction paraphrases and runtime restatements that preserve the goal.

Static shifts are applied at reset; runtime interventions are triggered at task-relevant stages. Task goals and success predicates remain unchanged. All variants of a base task stay in the same partition: 32 training base tasks and 8 held-out evaluation base tasks, with two held-out tasks per suite.

The split manifest selects 339 held-out instances (8 BASE + 331 shifted); the complementary catalogue has 1,349 training instances (32 BASE + 1,317 shifted). See catalogue details for task IDs and the mapping from service D20/D21 to paper D19/D20.

🛠️ Installation

Download evaluation.zip for the source, BDDL and 40 initial-state files, and assets.zip for simulator objects, scenes, textures and perturbation assets. The preparation script verifies both archives and every member before extracting a complete installation; no additional GitHub clone or Git LFS download is needed.

pip install -U huggingface_hub
hf download zhouxueyang/LIBERO-Pro2.0-Eval --repo-type dataset \
  --include "evaluation.zip" "evaluation.zip.sha256" "assets.zip" "assets.zip.sha256" \
            "manifests/**" "scripts/**" "docs/**" "README.md" "LICENSE_NOTICE.md" "release.json" \
  --local-dir ./LIBERO-Pro2.0-Eval

python ./LIBERO-Pro2.0-Eval/scripts/prepare_evaluation.py \
  --download-dir ./LIBERO-Pro2.0-Eval --output-dir ./libero-pro-install

Use a new output directory and reserve at least 3 GB for the release and extraction, plus dependencies, policy checkpoints and results. The reference environment requires Linux, Conda and compatible NVIDIA/EGL drivers. Install and check the environment:

cd libero-pro-install/LIBERO-Pro-evaluation
export LIBERO_PRO_CONDA_PREFIX="$PWD/.envs/libero-pro"
bash install.sh "$LIBERO_PRO_CONDA_PREFIX"
bash configure.sh "$LIBERO_PRO_CONDA_PREFIX"
bash start.sh
bash healthcheck.sh
set -a
source client.env
set +a
"$LIBERO_PRO_CONDA_PREFIX/bin/python" smoke_test.py

The extracted directory structure should look like:

libero-pro-install/
└── LIBERO-Pro-evaluation/
    ├── service-source/
    │   ├── libero_remote/
    │   ├── assets/
    │   │   ├── bddl/libero-pro2.0-1680-v1/
    │   │   ├── task_manifest.json
    │   │   └── benchmark_split_v1.json
    │   └── scripts/
    │       ├── run_grouped_eval.py
    │       └── run_policy.py
    ├── simulator-source/libero/libero/
    │   ├── assets/
    │   └── init_files/
    ├── requirements-lock.txt
    ├── install.sh
    ├── configure.sh
    └── start.sh

See the installation and evaluation guide for dependencies and deployment, and asset documentation for source provenance. The source and assets retain their original license notices. The smoke test checks the target simulator deployment; it is not a scored policy evaluation.

🔧 Evaluation

Run the policy in its official inference environment and connect it through an adapter with reset(context), act(observation) and close(). Install the client there with python -m pip install -e ./service-source, then load client.env. Obtain checkpoints and model dependencies from the corresponding model release; they are not included here.

From LIBERO-Pro-evaluation/, preview and submit a held-out run:

set -a
source client.env
set +a
python service-source/scripts/run_grouped_eval.py \
  --base-url "$LIBERO_REMOTE_URL" \
  plan --split test --group all --seed 142 --num-seeds 30 \
  --init-state-ids 0 --control-mode relative --output-dir results/plan

python service-source/scripts/run_grouped_eval.py \
  --base-url "$LIBERO_REMOTE_URL" \
  submit --split test --group all --seed 142 --num-seeds 30 \
  --init-state-ids 0 --control-mode relative --output-dir results/run

This example schedules 10,170 episodes (339 instances × 30 seeds, fixed initial-state ID 0); it does not assert the paper used this exact seed schedule. Use --group static or --group dynamic to evaluate those settings separately. Keep each model's official action horizon, preprocessing, control convention and suite-specific episode limits; use --max-steps for the official episode limit and change --control-mode when required by the policy.

Replace RUN_ID with the submitted run ID, and your_policy:create_adapter / policy_config.json with your own adapter factory and checkpoint configuration:

python service-source/scripts/run_policy.py \
  --base-url "$LIBERO_REMOTE_URL" --run-id RUN_ID \
  --adapter-factory your_policy:create_adapter \
  --adapter-kwargs policy_config.json --output-dir results/run

python service-source/scripts/run_grouped_eval.py \
  --base-url "$LIBERO_REMOTE_URL" export --run-id RUN_ID --output-dir results/run

The runner keeps the policy loaded across episodes. Resume interrupted evaluations with the same run_id; exports contain per-episode success in JSONL/CSV and grouped summaries. See the evaluation guide for the adapter contract, parallel runners and persistent results.

Citation

If you use LIBERO-Pro2.0, please acknowledge this project and cite the repository revision used:

@misc{zhou2026liberopro2,
  title = {LIBERO-Pro2.0: A Benchmark for Evaluating the Robustness of Embodied Foundation Models},
  author = {Xueyang Zhou and Yangming Xu and Qianjiang Li and Guiyao Tie and Junhao Wu and Ziheng Huang and Guowen Zhang and Zhijie Yuan and Di Wu and Duanfeng Chu and Pan Zhou and Lichao Sun and Yongchao Chen},
  year = {2026},
  howpublished = {GitHub repository},
  url = {https://github.com/Zxy-MLlab/LIBERO-Pro2.0}
}