---
license: apache-2.0
datasets:
- lerobot/libero
language:
- en
base_model:
- H-EmbodVis/TurboVLA
pipeline_tag: robotics
tags:
- ImageTextToAction
- Vision-Language-Action
- Robotics
- Multimodal
- Transformers
- Lightweight-VLA
- ViT
- Action-Prediction
---

![TurboVLA-architecture-diagram](https://cdn-uploads.huggingface.co/production/uploads/693830243222b0b00a4d04d1/cgJ_P6xfwXk1LJ40yxrMT.png)

## License & Attribution

This repository contains weights or code derived from the **TurboVLA** foundational architecture developed by **Hugging Face and the TurboVLA Authors**. 

* **License:** Distributed under the https://apache.org. 
* **Original Model Base:** https://huggingface.co/H-EmbodVis/TurboVLA
* **Copyright Notice:** Copyright 2025-2026 Hugging Face & the TurboVLA Authors.
* **Original Authors:** Hengyi Xie, Chenfei Yao, Xianjin Wu, Yingying Zhu, Dingkang Liang, Xiang Bai, Han Ding
* **Research Paper:** https://arxiv.org/abs/2607.27205

### Script to load this model:
```
import sys
import torch
import yaml
from pathlib import Path
from huggingface_hub import snapshot_download

# 1. Pull down your fully packaged repository tree to an isolated workspace cache
repo_id = "Man1103/TurboVLA-Libero-0.22B"
cache_dir = Path("./turbovla_cached_checkpoint")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

print(f"Streaming ready assets from {repo_id}...")
snapshot_download(repo_id=repo_id, local_dir=cache_dir)

# 2. Dynamically stitch the repository's native execution files into your Python context
# This allows you to load TurboVLA natively without manual git clones
sys.path.append(str(cache_dir))

# 3. Import the architecture templates included inside your repository package
from turbovla.models.turbovla import TurboVLAPolicy 

# 4. Parse your embedded structural configurations
with open(cache_dir / "config.yaml", "r") as f:
    config = yaml.safe_load(f)

# 5. Build the structural skeleton and map your ready weights directly onto your GPU
print("Assembling TurboVLA direct V+L -> A mapping blueprint...")
model = TurboVLAPolicy(config["model_config"]).to(device)

# Automatically match any .pth or weight binaries stored inside your repo
weight_file = list(cache_dir.glob("**/*.pth"))[0] 
checkpoint = torch.load(weight_file, map_location=device)

# Load the ready state dictionary safely into position
model.load_state_dict(checkpoint["model_state_dict"] if "model_state_dict" in checkpoint else checkpoint)
model.eval()

print(f"\n--- SUCCESS ---")
print(f"Your fully ready model is loaded onto {device} and primed for 32Hz LIBERO rollouts!")
```