OPENVLA Worklog

• Use pyproject.toml to set up OpenVLA (PyTorch versions are defined here).
• Install FlashAttention and BitsAndBytes versions compatible with the pyproject.toml.

Action Comparison Pipeline

1) Download the Bridge-v2 dataset

• Change directory to your base datasets folder:

cd <PATH TO BASE DATASETS DIR>

• Download the full dataset (approximately 124 GB):

wget -r -nH --cut-dirs=4 --reject="index.html*" https://rail.eecs.berkeley.edu/datasets/bridge_release/data/tfds/bridge_dataset/

• Rename the dataset to bridge_orig (required; omitting this may lead to runtime errors later):

mv bridge_dataset bridge_orig

2) Generate action-comparison JSONs

Run the following script:

python scripts/action-compare-pipeline/bridgev2_eval.py \
  --num_action_samples 8 \
  --sample_temperature 0.8

Notes

• Using quantized model variants works better with an RTX 4090.