User Guide · Policies · OpenVLA / OFT

OpenVLA & OpenVLA-OFT

MetaFine supports both the original OpenVLA 7B model and its OpenVLA-OFT (optimised fine-tuning) variant. Both consume the RLDS export, and both ship the same evaluate.py contract.

What they are

OpenVLA (7B vision-language-action model) and OpenVLA-OFT (its optimised fine-tuning variant). Vendored under core/policies/openvla and core/policies/openvla-oft.

Data — RLDS

Both consume the RLDS export. Produce it with utils/convert_to_rlds.py after recording (see Quickstart).

$ python utils/convert_to_rlds.py -i demos/lid_opening/100221 -o demos/datasets/rlds --dataset-name lid_opening_rlds --image-size 256

Install & run

$ pip install -e core/policies/openvla
$ pip install -e core/policies/openvla-oft

OpenVLA / OFT training + evaluation flags live in their own in-repo guides:

$ cat core/policies/openvla/README.md
$ cat core/policies/openvla-oft/README.md

Status: vendored. Paths are openvla and openvla-oft (hyphen). Treat the per-package READMEs as the source of truth.