Skip to content

Latest commit

 

History

History
46 lines (34 loc) · 1.27 KB

libero.md

File metadata and controls

46 lines (34 loc) · 1.27 KB

LIBERO Benchmark

LIBERO is a benchmark for studying knowledge transfer in multitask and lifelong robot learning problems.

Environment Setup

apt-get install libgl1-mesa-dri

cd reference/RoboVLMs

bash scripts/setup_libero.sh

Dataset download from huggingface.

Dataset Preparation

# 1. process the dataset, we process all the LIBERO suites together
python tools/process/libero_process.py

# 2. extract the vq tokens, need to change the dataset & output path, without augmentation
bash scripts/tokenizer/extract_vq_emu3.sh 

# 3. pickle generation for training
python tools/pickle_gen/pickle_generation_libero.py\
  --dataset_path ./datasets/processed_data \
  --output_path ./datasets/processed_data/meta \
  --normalizer_path ./configs/normalizer_libero \
  --output_filename libero_all_norm.pkl

Model Training

# default is one node training, recommend multi-node training.
bash scripts/simulator/libero/train_libero_video.sh

Model Evaluation

cd reference/RoboVLMs

# 1 GPU inference, modify the {task_suite_name} for different tasks
bash scripts/run_eval_libero_univla.sh ${CKPT_PATH}