fine-tuning-openvla-oft

Fine-tune and evaluate OpenVLA-OFT policies for robot action generation in LIBERO and ALOHA.

Updated May 2, 2026

Search Tags:#deployment #fine-tuning #robotics #lora #openvla #continuous actions #rlds

Ready to run

npx skills add https://github.com/qcmuu/AI-Research-Skills --skill fine-tuning-openvla-oft-qcmuu

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: fine-tuning-openvla-oft
Source: https://github.com/qcmuu/AI-Research-Skills/tree/main/18-multimodal/openvla-oft
Command: npx skills add https://github.com/qcmuu/AI-Research-Skills --skill fine-tuning-openvla-oft-qcmuu

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This Skill removes the complexity of reproducing and debugging OpenVLA-OFT and OpenVLA-OFT+ training and evaluation for robot action generation, so you can reliably match reported success rates in LIBERO simulation and ALOHA real-robot settings.

Core Features & Use Cases

OFT vs OFT+ configuration guidance: Select the correct continuous action heads, FiLM conditioning, and camera stream setup for LIBERO (OFT) versus ALOHA (OFT+).
End-to-end training and evaluation workflows: Run LIBERO evaluation on pretrained checkpoints and fine-tune with LoRA, including checkpoint selection and LoRA merging for deployment.
Production-oriented inference setup: Deploy a VLA server-client flow for ALOHA and verify critical config invariants (action head mode, image counts, FiLM, proprio, unnormalization keys).

Quick Start

Use the fine-tuning-openvla-oft skill to evaluate an OpenVLA-OFT checkpoint on LIBERO by running the provided LIBERO evaluation command with your chosen pretrained checkpoint and task suite.

Frequently Asked Questions about fine-tuning-openvla-oft

High-intent search queries and answers about installing and using this skill.

FAQPage Schema

How do I fine-tune OpenVLA-OFT for continuous robot actions in simulation?▼

To fine-tune OpenVLA-OFT for continuous robot actions, configure continuous action heads and use LoRA adaptation with consistent training flags, then run the provided evaluation workflows on the LIBERO simulation benchmark.

What is the difference between OFT and OFT+ configurations for OpenVLA robot policies?▼

OFT targets LIBERO simulation while OFT+ is built for ALOHA real-robot deployments, differing by specific continuous action heads, optional FiLM conditioning, and multi-camera input setups for robot action generation.

Can I use LoRA to fine-tune OpenVLA-OFT and merge checkpoints for deployment?▼

Yes, you can use LoRA adaptation to fine-tune OpenVLA-OFT policies, which includes checkpoint selection and LoRA merging to streamline production-oriented inference deployment for ALOHA server-client flows.

Why does my OpenVLA-OFT inference setup fail during ALOHA real-robot evaluation?▼

OpenVLA-OFT inference fails if critical config invariants are mismatched; you must verify consistent training and inference flags like use_l1_regression, use_film, num_images_in_input, lora_rank, and unnorm_key.

Do I need specific environment dependencies to run OpenVLA-OFT+ evaluation?▼

Yes, running OpenVLA-OFT+ evaluation requires setting up your environment with specific robotics, ML, and serving dependencies to support the VLA server-client flow and continuous action generation.