# TurboVLA — 0.2B robot policy runs at 32 Hz on RTX 4090 in under 1 GB VRAM

> A tiny vision-language-action model that keeps up with a real robot on a gaming GPU.

- Org: HUST + Huawei
- Date: 2026-07-29
- Category: paper, algorithm, repo
- Source: https://arxiv.org/abs/2607.27205
- Canonical page: https://ai-tldr.dev/releases/h-embodvis-turbovla/

## In short

TurboVLA is a 0.2B-parameter vision-language-action model that swaps the usual V→L→A pipeline for a direct V+L→A path, hitting 97.7% on LIBERO at 32 Hz on an RTX 4090 with under 1 GB VRAM.

## What it is

TurboVLA is an open-source vision-language-action model for robot manipulation from Huazhong University of Science and Technology and Huawei. The 0.2B-parameter policy runs at 32 Hz on a single RTX 4090 in under 1 GB of VRAM, and posts 97.7% average success on the LIBERO benchmark.

## How it works

Most VLAs route the camera through a large language model before predicting an action (V→L→A). TurboVLA drops that path and encodes vision and language in parallel, exchanges information through a lightweight bidirectional cross-attention block, and predicts action chunks with a compact decoder. Inference is 31.2 ms per step, so the loop keeps up with a 30 Hz control rate on a consumer GPU.

## Why it matters

Real-time robot control usually needs a workstation-grade GPU or a distilled policy. TurboVLA runs the full policy in real time on a laptop-class card with room to spare, which lowers the bar for hobbyist and university robotics groups that already have a 40-series GPU. The paper also reports 60.2% on RoboTwin 2.0 bimanual tasks.

## Who it's for

Robotics researchers and hobbyists running LIBERO / RoboTwin on a single RTX-40 GPU.

## Try it

git clone https://github.com/H-EmbodVis/TurboVLA

## Links

- [arXiv paper](https://arxiv.org/abs/2607.27205)
- [Code (GitHub)](https://github.com/H-EmbodVis/TurboVLA)
- [Project page](https://h-embodvis.github.io/TurboVLA/)
- [PDF](https://arxiv.org/pdf/2607.27205)