OpenVLA 및 LIBERO 환경 구축

 # OpenVLA 및 LIBERO 환경 구축

- [참고] https://kimelab.notion.site/OpenVLA-LIBERO-24f875de525f813984b9f84b75e5d5fb

- [다운로드]

(1) finetuning version: https://huggingface.co/openvla/openvla-7b-finetuned-libero-spatial/tree/main 의 모든 파일 다운로드

(2) original: https://huggingface.co/openvla/openvla-7b/tree/main

- 연구실의 우분투 서버: /home/vislab/.cache/huggingface/hub/openvla-7b-finetuned-libero-spatial/

            폴더에 학습된 pt파일 및 관련 설정파일, 소스 등이 저장되어 있음

- [실행] ~/Downloads/openvla위치로 가서 

    > xvfb-run -a python experiments/robot/libero/run_libero_eval.py --model_family openvla --pretrained_checkpoint    /home/vislab/.cache/huggingface/hub/openvla-7b-finetuned-libero-spatial/ --task_suite_name libero_spatial --center_crop True --num_trials_per_task 1

- [확인] /home/vislab/Downloads/openvla/rollouts/2026_02_26/ 폴더 위치에 만들어 지는 mp4파일을 열어봄

## Set-up: Conda Env

# Create and activate conda environment

conda create -n openvla python=3.10 -y

conda activate openvla

# CUDA 12.1 (Conda)

conda install pytorch==2.2.0 torchvision==0.17.0 torchaudio==2.2.0 pytorch-cuda=12.1 -c pytorch -c nvidia -y 

## Set-up: Openvla Repo

# Clone and install the openvla repo

git clone https://github.com/openvla/openvla.git

cd openvla

pip install -e .

# openvla pip 설치이후 Pytorch CPU 버전으로 downgrade 현상 체크

## Set-up(for Ubuntu): Ninja & Flash Attention 2

# Install Flash Attention 2 for training (https://github.com/Dao-AILab/flash-attention)

#   =>> If you run into difficulty, try `pip cache remove flash_attn` first

pip install packaging ninja

# ninja --version; echo $?  # Verify Ninja --> should return exit code "0"

# Linux GCC 필수 (윈도우에서는 MSVC 필요) 

pip install "flash-attn==2.5.5" --no-build-isolation

## Inference Test — OpenVLA (Linux & Windows)

- [실행 스크립트] Source code: vla-scripts/extern/verify_openvla.py

# Install — LIBERO

## Clone and install the LIBERO repo

git clone https://github.com/Lifelong-Robot-Learning/LIBERO.git

cd LIBERO

pip install -e .

cd ..

## Install other required packages for OpenVLA

cd openvla

pip install -r experiments/robot/libero/libero_requirements.txt

# Launching LIBERO Evaluations (Only Linux)

## Set-up — 4-bit 양자화 모드

`openvla/experiments/robot/libero` 폴더 > `run_libero_eval.py` 

GenerateConfig 클래스 > `load_in_4bit: bool = False` → `load_in_4bit: bool = True`

# Launch LIBERO-Spatial evals

python experiments/robot/libero/run_libero_eval.py \

  --model_family openvla \

  --pretrained_checkpoint openvla/openvla-7b-finetuned-libero-spatial \

  --task_suite_name libero_spatial \

  --center_crop True

  

python experiments/robot/libero/run_libero_eval.py --model_family openvla --pretrained_checkpoint openvla/openvla-7b-finetuned-libero-spatial --task_suite_name libero_spatial --center_crop True

# Launch LIBERO-Object evals

python experiments/robot/libero/run_libero_eval.py \

  --model_family openvla \

  --pretrained_checkpoint openvla/openvla-7b-finetuned-libero-object \

  --task_suite_name libero_object \

  --center_crop True

# Launch LIBERO-Goal evals

python experiments/robot/libero/run_libero_eval.py \

  --model_family openvla \

  --pretrained_checkpoint openvla/openvla-7b-finetuned-libero-goal \

  --task_suite_name libero_goal \

  --center_crop True

# Launch LIBERO-10 (LIBERO-Long) evals

python experiments/robot/libero/run_libero_eval.py \

  --model_family openvla \

  --pretrained_checkpoint openvla/openvla-7b-finetuned-libero-10 \

  --task_suite_name libero_10 \

  --center_crop True