Run and evaluate Hugging Face’s LeRobot policies on the LIBERO benchmark for lifelong robot learning.

LeRobot-LIBERO bridges Hugging Face's LeRobot and LIBERO, enabling evaluation of vision-language-actions (VLAs) models and imitation learning policies on standardized robotic manipulation tasks. This repository supports policy inference, reproducible evaluation, and integration with LIBERO's task suites.

git clone [email protected]:JiahongChen/lerobot-libero.git
cd lerobot-libero

conda create -y -n lerobot-libero python=3.10
conda activate lerobot-libero

git clone https://github.com/huggingface/lerobot.git
cd lerobot
conda install -y ffmpeg -c conda-forge
pip install -e .
pip install -e ".[smolvla, pi0]"
cd ..

git clone https://github.com/Lifelong-Robot-Learning/LIBERO.git
cd LIBERO
pip install -e .
cd ..

pip install -r libero_requirements.txt

Use the following command to run inference on a selected LIBERO task suite:

python lerobot_inference.py \
    --policy_path peeeeeter/smolvla_spatial \
    --task_suite_name [libero_spatial | libero_object | libero_goal | libero_10 | libero_90]

Replace --policy_path with your desired pretrained policy and select an appropriate --task_suite_name.

• lerobot_inference.py: Entry point for evaluating LeRobot models on LIBERO tasks.
• libero_requirements.txt: Dependencies for running LIBERO tasks with LeRobot.
• Other files coming soon...

• Ensure you are using Python 3.10 for compatibility with both LeRobot and LIBERO.
• ffmpeg is required for video-related operations and is installed via conda.
• The script supports evaluating pretrained models hosted on Hugging Face Hub or locally.

• LeRobot: Hugging Face's large behavior model framework.
• LIBERO: A benchmark suite for lifelong robot learning.