Skip to content
Get packs

Fine-Tuning OpenVLA-OFT

Guides fine-tuning and evaluating OpenVLA-OFT robot action models on LIBERO and ALOHA setups

Source Download skill
Repository
Orchestra-Research/AI-Research-SKILLs
Stars
14,195
Licence
MIT
Files
5 files

What it does

This skill gives an AI agent a detailed playbook for OpenVLA-OFT, a research method for training robot control models built on the OpenVLA vision-language model. Once installed, the agent can walk you through setting up the exact software environment, running evaluation on pretrained LIBERO simulation checkpoints, and following fine-tuning and deployment steps for LIBERO and ALOHA robot setups. It also helps debug common problems such as normalization, LoRA merging, and cross-GPU issues.

Great for

  • Researchers trying to reproduce the OpenVLA-OFT paper results on LIBERO or ALOHA
  • Robotics engineers training custom action heads on top of OpenVLA with LoRA
  • Teams deploying a server-client inference setup for ALOHA robots

Skip it if

  • Users without access to a high-end GPU setup (fine-tuning assumes 8x A100 class hardware and evaluation needs roughly 16 GB+ VRAM), since the workflows will not run on a typical laptop
  • People who want general chatbot or LLM fine-tuning, which this skill explicitly does not cover
  • Anyone expecting to run real robots without safety review, since the skill covers training and evaluation workflows and real-robot deployment carries physical risk

Try asking

  • “Set up a conda environment with the exact PyTorch 2.2.0 and PEFT 0.11.1 versions needed for OpenVLA-OFT”
  • “Evaluate the moojink LIBERO-Spatial checkpoint with 50 trials per task and seed 7”
  • “Parse a LIBERO evaluation log and print per-task success rates”
  • “Help me debug a LoRA merge problem after upgrading torch”
Step-by-step guide: how to use Fine-Tuning OpenVLA-OFTInstall, first run, examples, tips and FAQs

Install in 3 steps

  1. 1

    Download the skill

    The ZIP contains the 5 original files, the MIT licence and an attribution note.

    openvla-oft.zip
  2. 2

    Add it to your agent

    Claude Code (all projects)

    mkdir -p ~/.claude/skills && unzip openvla-oft.zip -d ~/.claude/skills/

    Claude Code / Agent SDK (this project only)

    mkdir -p .claude/skills && unzip openvla-oft.zip -d .claude/skills/

    Claude apps (web/desktop): Settings → Capabilities → Skills → Upload skill, and choose the ZIP.

  3. 3

    Ask for the task

    No special command needed — the agent reads the skill's description and loads it when your request matches. Try one of the prompts above.

Get better results

  • Pin the exact versions listed in the skill (torch 2.2.0, peft 0.11.1); the author warns that upgrading torch may break the LoRA merge and action head outputs
  • Start with the pretrained LIBERO evaluation before attempting fine-tuning, since it is far cheaper and confirms your setup works
  • Review any commands the agent proposes before running them, since the workflows clone repos, install packages, and launch long GPU jobs

What's inside

  • references/aloha-workflow.md
  • references/config-troubleshooting.md
  • references/libero-workflow.md
  • references/paper-and-checkpoints.md
  • SKILL.md

Redistributed unmodified from Orchestra-Research/AI-Research-SKILLs under the MIT licence. The assessment was drafted with AI from the skill's files on 2026-10-29. Skills can instruct your agent to run code — read SKILL.md and any scripts before installing.