Libero - Train Your Own Model

This section describes how to train your own model using Libero data and configurations.

Prerequisites

• Complete the environment installation (refer to :doc:installation/index).

• Complete data preparation (refer to :doc:data/index).

• Prepare a training configuration file and pre-trained weights (optional).

Quick Start

We use single-node multi-GPU fine-tuning of the Pi0.5 model as an example. First, prepare the Pi0.5 pre-trained model weights, or download them from HuggingFace.

You also need to prepare the LIBERO-10 dataset in LeRobot format. Refer to the data preparation documentation for download instructions. Then set the environment variables. Adjust MLP_WORKER_GPU according to the actual number of GPUs:

export MLP_WORKER_GPU=8
export MLP_WORKER_NUM=1
export MLP_ROLE_INDEX=0
export MLP_WORKER_0_HOST=localhost
export MLP_WORKER_0_PORT=29500

Modify the dataset path in configs/pi05/pi05_paligemma_libero10_full_finetune.py:

    ...
    freeze_vision_backbone=False,
    pretrained_name_or_path=  # noqa: E251
    './checkpoints/pi05_libero/model.safetensors',  # Replace with the actual pre-trained weight path
    name_mapping={
            ...

    ...
    datasets=dict(
            type='ParquetDataset',
            data_root_path=  # noqa: E251
            './datasets/libero_10_no_noops_lerobotv2.1',  # Replace with the actual dataset path
            transforms=[
                ...

Launch training using scripts/train.sh:

bash scripts/train.sh \
    configs/pi05/pi05_paligemma_libero10_full_finetune.py \
    work_dirs/pi05_paligemma_libero10_full_finetune # Replace with your desired output directory

Training Artifacts

After training, the work_dirs/... directory typically contains:

• checkpoint_*.pt model files

• Logs and configuration backup files

For more training parameter details, refer to :doc:vla.