VLA-Adapter

AI model
Humanoids and robot learning

A vision-language-action model that adds a lightweight adapter bridge to a small Qwen2.5-0.5B backbone so it can learn robot control; the code is fine-tuned and evaluated on LIBERO and CALVIN.

Made by OpenHelix Robotics.

Source on GitHub: github.com/OpenHelix-Team/VLA-Adapter (opens in a new tab). Facts read at commit 23fa0c9 (opens in a new tab).

Licenses

Code: MIT (opens in a new tab)
Read by Robotheca from LICENSE (opens in a new tab) (MIT) at commit 23fa0c9 (opens in a new tab), on 16 Oct 2026.
Weights: MIT (opens in a new tab)
Read on huggingface.co/VLA-Adapter/LIBERO-Long (opens in a new tab).
Commercial use: Yes

What it does

Fine-tunes a Prismatic VLM built on Qwen2.5-0.5B using an adapter bridge paradigm for robot action prediction.
Provides training configurations for GPUs from roughly 10GB to 80GB of VRAM, with LoRA and batch-size tuning.
Released checkpoints for LIBERO Spatial, Object, Goal and Long suites, plus a Pro version and a CALVIN ABC checkpoint.
Includes evaluation scripts for LIBERO and CALVIN, with inference throughput measurement suggested via timing code.
Recently added real-world ALOHA deployment support, verified on the Cobot Magic robot.
Planned work includes support for Franka, UR-5 and AGILE Piper systems, though these are listed as TODO items.

Written by Robotheca from the project's README (opens in a new tab) at commit 23fa0c9.

Getting started

Difficulty: Advanced
Hardware it needs, as the project states it: Flash Attention 2 is required for training. The ALOHA real-world deployment runs on a Cobot Magic robot.
How to start: The project's website (opens in a new tab)

Maintenance

Status: Maintained: a commit in the last 12 months
Newest commit: 19 Mar 2026 on main
Latest release: None published on GitHub

Checked on GitHub on 16 Oct 2026.

Robots on Robotheca that use it

No robot on Robotheca is linked to it yet. Links come from a robot's robot.yaml, from its dependency files read as text, or from Robotheca's editors.

Robotheca lists AI models as a reference: nothing is hosted or run here, and every fact links to where it was read.