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.

This repository contains the official implementation of TurboVLA for the paper TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM.

_{Real-world tasks with synchronous policy inference.}

---

📣 News

• 2026.09.27:🚀TurboVLA achieves an 88.06% average success rate on RoboTwin 2.0. The paper, code, and checkpoints have been updated.
• 2026.07.31: Released the TurboVLA model checkpoints on Hugging Face.
• 2026.07.30: Released the paper, training and evaluation code.

📅 TODO

• Support Huawei Ascend NPUs

---

📄 Abstract

Vision-language-action (VLA) models commonly adopt an LLM-centric $V\rightarrow \boldsymbol{L}\rightarrow A$ pathway, processing visual observations and language instructions through a large language model before predicting robot actions. Although effective, this design incurs substantial computation and memory overhead.

In this work, we introduce TurboVLA, a compact VLA architecture built on a direct $\boldsymbol{V}+\boldsymbol{L}\rightarrow A$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 directly constructs task-conditioned representations while avoiding the overhead of an LLM-centric execution pathway. On LIBERO, TurboVLA achieves 97.6% average success with only 0.2B parameters, 31.2ms inference latency, and 0.9GiB inference VRAM on a consumer-grade RTX 4090. Notably, a 0.4B TurboVLA achieves 88.06% success on RoboTwin 2.0, even matching or outperforming substantially larger VLA policies. These results demonstrate that the simple $\boldsymbol{V}+\boldsymbol{L}\rightarrow A$ design of TurboVLA can achieve high performance without requiring an LLM-centric execution pathway, offering a new perspective on how vision, language, and action can be connected for efficient robotic manipulation.

---

🔍 Overview

---

📈 Performance

---

⚙️ Installation

Clone the repository:

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

LIBERO and RoboTwin use different simulator and data stacks. We recommend separate Python 3.10 environments.

LIBERO Environment

conda create -n turbovla-libero python=3.10 -y
conda activate turbovla-libero

# Install the CUDA-compatible PyTorch build for your system first.
pip install torch==2.3.1 torchvision==0.18.1 --index-url https://download.pytorch.org/whl/cu121
pip install -e ".[libero]"

Install LIBERO separately in the same environment.

RoboTwin Environment

conda create -n turbovla-robotwin python=3.10 -y
conda activate turbovla-robotwin

# Install a CUDA-compatible PyTorch build before the project dependencies.
pip install -e ".[robotwin]"
pip install flash-attn==2.7.4.post1 --no-build-isolation

Install the RoboTwin 2.0 simulator in a separate evaluation environment when required by your setup.

---

📦 Dataset and Model Preparation

Model weights and benchmark datasets are external assets and are not committed to this repository.

Pretrained Checkpoints

Download the complete TurboVLA release, including model weights and normalization metadata, from Hugging Face:

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

Required Models

Asset	Source	Used by
DINOv3 ViT-B	facebookresearch/dinov3	LIBERO
DINOv3 ViT-L	facebookresearch/dinov3	RoboTwin
BERT base uncased	google-research/bert	Both
GroundingDINO Swin-T OGC	IDEA-Research/GroundingDINO	Both

LIBERO Data

TurboVLA expects the four modified no-noops suites in TFDS/RLDS format:

data/libero/
|-- libero_10_no_noops/1.0.0/
|-- libero_goal_no_noops/1.0.0/
|-- libero_object_no_noops/1.0.0/
`-- libero_spatial_no_noops/1.0.0/

The repository provides no-op removal and mixed-suite statistics utilities:

python scripts/libero/regenerate_libero_no_noops.py --help
python scripts/libero/compute_mixed_stats.py --help

Released normalization statistics are stored in experiments/libero/configs/libero_all4_stats.json. BERT is part of the model and runs online during both training and evaluation; no text-feature cache is required. See experiments/libero/README.md for details.

RoboTwin Data

Download both converted LeRobot variants into one root. The script downloads StarVLA/RoboTwin-Clean and the Randomized/ tree from StarVLA/RoboTwin-Randomized, then checks all 50 task pairs:

bash scripts/robotwin/prepare_data.sh /path/to/converted/RoboTwin
export ROBOTWIN_DATA_ROOT=/path/to/converted/RoboTwin

Install the Hugging Face hf CLI first. The resulting layout is Clean/<task_name> and Randomized/<task_name>. See experiments/robotwin/README.md for details.

---

🏋️ Training and Evaluation

LIBERO Training

The paper recipe uses DINOv3 ViT-B, two camera views, 7-D actions, a 12-step action chunk, 40k optimizer steps, 10k warmup steps, and global batch size 128 on four GPUs (per-device batch size 8 with 4 gradient-accumulation steps).

torchrun --nproc_per_node=4 experiments/libero/train.py \
  --dataset_dir data/libero/libero_10_no_noops/1.0.0 \
  --dataset_dirs "data/libero/libero_10_no_noops/1.0.0,data/libero/libero_goal_no_noops/1.0.0,data/libero/libero_object_no_noops/1.0.0,data/libero/libero_spatial_no_noops/1.0.0" \
  --stats_path experiments/libero/configs/libero_all4_stats.json \
  --stats_key libero_all4_no_noops \
  --dinov3_path facebook/dinov3-vitb16-pretrain-lvd1689m \
  --bert_path google-bert/bert-base-uncased \
  --allow_hf_download \
  --pretrained_init_ckpt /path/to/groundingdino_swint_ogc.pth \
  --batch_size 8 \
  --checkpoint_dir outputs/libero

LIBERO Evaluation

One command evaluates one checkpoint on one suite.

python experiments/libero/evaluate.py \
  --ckpt_path pretrained/TurboVLA/checkpoints/libero/libero_object.pth \
  --dinov3_path /path/to/dinov3-vitb \
  --bert_path /path/to/bert-base-uncased \
  --stats_path experiments/libero/configs/libero_all4_stats.json \
  --stats_key libero_all4_no_noops \
  --task_suite_name libero_object \
  --num_trials_per_task 50 \
  --chunk_size 12 \
  --num_open_loop_steps 12 \
  --seed 7 \
  --precision bf16 \
  --result_json_path outputs/evaluation/libero_object.json

Valid suite names are libero_spatial, libero_object, libero_goal, and libero_10.

RoboTwin Training

The RoboTwin recipe reported in the paper uses DINOv3 ViT-L, three camera views at 224×224, frozen BERT, 14-D absolute joint-position actions, and a 50-step ACT head. Tasks are sampled uniformly with Clean:Randomized at 1:10. Training uses eight GPUs × 64 samples (global batch 512), 150k optimizer steps, and a 5e-5 learning rate.

export ROBOTWIN_DATA_ROOT=/path/to/converted/RoboTwin
export BERT_MODEL_PATH=/path/to/bert-base-uncased
export TURBOVLA_INIT_CKPT=/path/to/groundingdino_swint_ogc.pth
export DINOV3_MODEL_PATH=/path/to/dinov3-vitl
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

RUN_ID=turbovla_robotwin_all50_taskbalanced_150k \
bash scripts/robotwin/train.sh

RoboTwin Evaluation

The policy server and RoboTwin simulator can run in separate Python environments:

export ROBOTWIN_PATH=/path/to/RoboTwin
export STARVLA_PYTHON=/path/to/policy-env/bin/python
export ROBOTWIN_PYTHON=/path/to/robotwin-env/bin/python
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

bash scripts/robotwin/evaluate.sh \
  /path/to/steps_150000_ema_pytorch_model.pt

By default, the command evaluates all 50 tasks in both Clean and Randomized. Use --mode clean or --mode randomized for one variant, and append task names for a subset. Check out RoboTwin 2.0 commit bf44be5 for this evaluation adapter. That simulator version evaluates 100 episodes per task; its eval_policy.py does not read ROBOTWIN_TEST_NUM.

---

👍 Acknowledgement

TurboVLA builds upon the following projects and resources:

• DINOv3 for visual representations.
• GroundingDINO for bidirectional vision-language interaction components and initialization.
• VLA-Adapter for the LIBERO task and episode rollout protocol.
• StarVLA for the RoboTwin-compatible training and evaluation runtime.
• LIBERO and RoboTwin 2.0 for simulation benchmarks.

---

📖 Citation

If TurboVLA is useful in your research, please consider citing the paper:

@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}
}