| --- |
| language: |
| - en |
| tags: |
| - gguf |
| - robotics |
| - embodied-ai |
| - vision-language-action |
| - world-action-model |
| - cpp |
| - inference |
| library_name: gguf |
| pipeline_tag: robotics |
| license: other |
| --- |
| |
| # Embodied.cpp GGUF Models |
|
|
| This repository provides **GGUF model weights converted for use with [Embodied.cpp](https://arxiv.org/abs/2607.02501)**, a portable C++ inference runtime for embodied AI models on heterogeneous robots. |
|
|
| These files are **deployment-oriented conversions** of upstream checkpoints. They are intended for **inference with Embodied.cpp**, rather than for training or fine-tuning in the original Python stacks. |
|
|
| ## About Embodied.cpp |
|
|
| **Embodied.cpp** is a portable inference runtime for embodied AI models, including Vision-Language-Action (VLA) models and World-Action Models (WAMs). It focuses on low-latency closed-loop deployment across heterogeneous hardware and robot systems. |
|
|
| Paper: |
| - **Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots** |
| - arXiv: https://arxiv.org/abs/2607.02501 |
|
|
| ## What This Repository Contains |
|
|
| This repository hosts **GGUF-converted model artifacts** used by `embodied.cpp`. |
|
|
| The goal is to make upstream embodied model checkpoints easier to deploy from C++ with a unified runtime and serving interface. |
|
|
| ## Conversion Provenance |
|
|
| The GGUF files in this repository are converted from upstream checkpoints using the conversion scripts in the `embodied.cpp` project. |
|
|
| ### 1. pi0.5 |
|
|
| Upstream checkpoint: |
| - https://huggingface.co/lerobot/pi05_libero_finetuned_v044 |
| |
| Conversion scripts: |
| - `scripts/convert_pi05_to_gguf.py` |
| - `scripts/convert_pi05_mmproj_to_gguf.py` |
|
|
| Notes: |
| - The converted artifacts are intended for `embodied.cpp` pi0.5 inference. |
| - pi0.5 uses separate GGUF exports for the main policy weights and the multimodal projector / vision-side component. |
|
|
| ### 2. Hunyuan-VLA-0.5 |
|
|
| Upstream checkpoint: |
| - https://huggingface.co/tencent/Hy-Embodied-0.5-VLA-RoboTwin |
|
|
| Conversion code: |
| - Conversion scripts are included in the `embodied.cpp` repository. |
|
|
| Notes: |
| - These GGUF files target `embodied.cpp` deployment of HY-VLA on RoboTwin-style evaluation and related runtime paths. |
|
|
| ### 3. LingBot-VA |
|
|
| Upstream source: |
| - https://huggingface.co/datasets/robbyant/libero-long-lerobot |
|
|
| Conversion code: |
| - Conversion scripts are included in the `embodied.cpp` repository. |
|
|
| Notes: |
| - These GGUF artifacts are prepared for the LingBot-VA path in `embodied.cpp`. |
|
|
| ## Intended Use |
|
|
| These GGUF files are intended for: |
|
|
| - Running embodied AI models with `embodied.cpp` |
| - C++ inference and server deployment |
| - Simulation and robotics evaluation pipelines supported by `embodied.cpp` |
|
|
| These files are **not** the original training checkpoints and may not be directly compatible with upstream Python inference code without the `embodied.cpp` runtime. |
|
|
| ## Usage |
|
|
| Please use these weights with the corresponding `embodied.cpp` model server or evaluation pipeline. |
|
|
| Typical usage includes: |
|
|
| - `vla-pi05-server` |
| - `vla-hy-vla-server` |
| - `wam-lingbot-server` |
|
|
| Please refer to the `embodied.cpp` project repository and documentation for build instructions, model-specific runtime requirements, and evaluation examples. |
|
|
| ## Roadmap |
|
|
| This repository will continue to be updated with GGUF releases for: |
|
|
| - **pi0.5** |
| - **Hunyuan-VLA-0.5** |
| - **LingBot-VA** |
|
|
| Additional `embodied.cpp`-compatible model exports may be added in the future. |
|
|
| ## Attribution |
|
|
| All model credit belongs to the original upstream authors and maintainers. |
|
|
| Upstream sources referenced here: |
|
|
| - pi0.5 (LeRobot checkpoint): https://huggingface.co/lerobot/pi05_libero_finetuned_v044 |
| - HY-VLA RoboTwin checkpoint: https://huggingface.co/tencent/Hy-Embodied-0.5-VLA-RoboTwin |
| - LingBot-VA related source dataset: https://huggingface.co/datasets/robbyant/libero-long-lerobot |
| |
| ## License and Restrictions |
| |
| This repository contains converted deployment artifacts derived from upstream resources. |
| |
| Please review and comply with the **original upstream licenses, model terms, and dataset restrictions** before use or redistribution. |
| |
| Because different upstream sources may have different licenses or usage constraints, users are responsible for checking the original source pages. |
| |
| ## Citation |
| |
| If you find these GGUF conversions or `embodied.cpp` useful, please cite: |
| |
| ```bibtex |
| @article{xu2026embodiedcpp, |
| title={Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots}, |
| author={Xu, Ling and Han, Chuyu and Li, Borui and Wu, Hao and Jiang, Shiqi and Cao, Ting and Li, Chuanyou and Zhong, Sheng and Wang, Shuai}, |
| journal={arXiv preprint arXiv:2607.02501}, |
| year={2026} |
| } |
