Fine Tuning Openvla Oft

This wraps the official OpenVLA-OFT fine-tuning workflows for adapting vision-language-action models to robot tasks. You're getting LoRA-based fine-tuning pipelines and eval scripts for the LIBERO benchmark, which tests manipulation skills across different task suites. The quick start points you straight at evaluating a pretrained checkpoint on LIBERO spatial tasks with 50 trials per task. It's built for researchers doing embodied AI work who need to adapt OpenVLA's continuous action prediction for their own robot setups. The repo setup assumes you're comfortable cloning research codebases and running Python eval scripts. Three security audits show mostly green with one Snyk warning to check.

skills.sh installs: 926
GitHub stars: 14.4K
last updated: Aug 21, 2026

View on GitHub

Install

$ npx -y skills add orchestra-research/ai-research-skills --skill fine-tuning-openvla-oft --agent claude-code

Installs into .claude/skills of the current project.

Files

OpenVLA-OFT

Fine-tuning and evaluation workflows for OpenVLA-OFT and OpenVLA-OFT+ from the official openvla-oft codebase. Covers blank-machine setup plus LoRA-based adaptation of OpenVLA for robot action generation with continuous action prediction heads.

Quick start

Clone the public repo, follow the official setup, then evaluate a pretrained LIBERO checkpoint:

git clone https://github.com/moojink/openvla-oft.git
cd openvla-oft
python experiments/robot/libero/run_libero_eval.py \
  --pretrained_checkpoint moojink/openvla-7b-oft-finetuned-libero-spatial \
  --task_suite_name libero_spatial \
  --center_crop True \
  --num_trials_per_task 50 \
  --seed 7

Core concepts

What OpenVLA-OFT changes: Standard OpenVLA tokenizes continuous actions into discrete bins, losing precision. OFT replaces this with dedicated continuous action heads (L1 regression or diffusion) while keeping the VLA backbone frozen and adapting via LoRA.

OFT vs OFT+ variants: