TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM

Hengyi Xie^1*, Chenfei Yao^1*, Xianjin Wu¹, Xuanyang Xi², Yiping Tang², Di Xu², Yingying Zhu¹, Dingkang Liang^1†, Xiang Bai¹, Han Ding¹

¹ Huazhong University of Science and Technology, China
² Huawei Technologies Co. Ltd, China
^* Equal contribution, listed alphabetically by surname. ^† Project lead.

🔍 Overview

TurboVLA is a real-time vision-language-action model for robotic manipulation. It replaces the conventional LLM-centric vision-to-language-to-action pathway with a direct vision-plus-language-to-action design: visual observations and language instructions are encoded independently, exchanged through lightweight bidirectional interaction, and decoded into continuous action chunks.

This repository contains the official TurboVLA checkpoints for the LIBERO and RoboTwin 2.0 evaluations.

Abstract

Vision-language-action (VLA) models commonly adopt an LLM-centric V → L → A pathway, where visual observations are projected into the representation space of a large language model before being decoded into robot actions. Although effective, this design incurs substantial computation and memory overhead at every policy invocation. In this work, we introduce TurboVLA, a new VLA paradigm that reformulates the conventional V → L → A pathway as a direct V + L → A mapping. Instead of using a large language model as the central interface between perception and action, TurboVLA independently encodes visual observations and language instructions, directly exchanges information between them through lightweight bidirectional vision-language interaction, and predicts continuous action chunks with a compact decoder. This simple design constructs task-conditioned representations directly from visual and linguistic features, significantly reducing the computational and memory costs of VLA inference. On LIBERO, TurboVLA achieves 97.7% average success with only 0.2B parameters, 31.2 ms inference latency, and 0.9 GB inference VRAM on a consumer-grade RTX 4090, matching or outperforming substantially larger VLA policies. These results establish TurboVLA as a simple and effective alternative to the prevailing LLM-centric VLA paradigm, offering a new perspective on how vision, language, and action can be connected for efficient robotic manipulation.

✨ Key Ideas

• Direct vision-plus-language-to-action modeling without an LLM-centric action pathway.
• Independent DINOv3 visual and BERT language encoders with lightweight bidirectional interaction.
• Compact action-chunk decoders for low-latency and memory-efficient inference.
• Official checkpoints for both LIBERO and RoboTwin 2.0.

🎯 Task, Inputs, and Outputs

TurboVLA predicts continuous robot action chunks from synchronized multi-view RGB observations, a natural-language instruction, and the robot proprioceptive state.

Benchmark	Inputs	Output	Action horizon	Visual encoder
LIBERO	Two RGB views, language instruction, robot state	7-D continuous actions	12	DINOv3 ViT-B
RoboTwin 2.0	Three 224 x 224 RGB views, language instruction, 14-D robot state	14-D absolute joint-position actions	50	DINOv3 ViT-L

📦 Released Checkpoints

LIBERO

The LIBERO model was jointly trained on all four suites in a single mixed-suite training run. The released evaluation uses the following suite-specific checkpoints:

Suite	Evaluation name	Checkpoint
LIBERO-Object	libero_object	checkpoints/libero/object.pth
LIBERO-Goal	libero_goal	checkpoints/libero/goal.pth
LIBERO-Spatial	libero_spatial	checkpoints/libero/spatial.pth
LIBERO-Long	libero_10	checkpoints/libero/long.pth

The LIBERO files are model-state-only PyTorch checkpoints. Optimizer and scheduler states were removed; the released model_state_dict tensors are unchanged and are compatible with the official TurboVLA loader.

RoboTwin 2.0

Training data	Weights	Checkpoint
RoboTwin Clean, 50 tasks	EMA	checkpoints/robotwin/

🛠️ Usage

You can download the complete model repository using either the Hugging Face CLI or the Python SDK.

Option 1: Hugging Face CLI (Recommended)

pip install -U huggingface_hub
hf download H-EmbodVis/TurboVLA --local-dir TurboVLA-model

Option 2: Python SDK

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="H-EmbodVis/TurboVLA",
    repo_type="model",
    local_dir="TurboVLA-model",
)

TurboVLA is a custom PyTorch VLA implementation and is not directly loadable with transformers.AutoModel.from_pretrained(). For installation and evaluation instructions, use the official GitHub repository.

📂 Repository Contents

Path	Description
checkpoints/libero/*.pth	Four LIBERO suite-specific evaluation checkpoints
checkpoints/robotwin/*.safetensors	RoboTwin 2.0 EMA checkpoint
libero_all4_stats.json	Mixed-suite LIBERO normalization statistics
config.yaml	RoboTwin model and inference configuration
dataset_statistics.json	RoboTwin state and action normalization statistics
config.json	Hugging Face model metadata and checkpoint index
CHECKSUMS.sha256	SHA-256 checksums for all released checkpoints
DINOv3_LICENSE.md	DINOv3 terms applicable to the released checkpoint parameters
LICENSE	Apache-2.0 license for the TurboVLA project code

⚖️ License

The TurboVLA source code is released under Apache-2.0. The released checkpoints contain parameters derived from DINOv3 and are distributed subject to the included DINOv3 License. Users must comply with all applicable third-party license terms.

📖 Citation

If you find this work useful, please consider citing:

@article{xie2026turbovla,
  title  = {TurboVLA: Real-Time Vision-Language-Action Model at
            32 Hz on an RTX 4090 with <1 GB VRAM},
  author = {Xie, Hengyi and Yao, Chenfei and Wu, Xianjin and
            Xi, Xuanyang and Tang, Yiping and Xu, Di and
            Zhu, Yingying and Liang, Dingkang and Bai, Xiang and
            Ding, Han},
  journal = {arXiv preprint arXiv:2607.27205},
  year   = {2026}
}