Embodied.cpp

Thu, 02 Jul 2026 17:58:28 +0000

Embodied.cpp is a portable C++ inference runtime for embodied AI models, including vision-language-action (VLA) and world-action models. It provides a shared deployment path across heterogeneous CPUs, GPUs, NPUs, robots, and simulators without requiring a separate model-specific Python stack for every model.

Overview

Embodied AI models combine multimodal perception, reasoning, and control, but their deployment is often fragmented across Python environments, model-specific runtime code, and hardware-specific execution paths. Embodied.cpp turns this into a unified systems stack: it converts supported models to GGUF, exposes a common runtime architecture, and provides reusable serving and evaluation paths.

Project leads: Borui Li and Hao Wu.

Highlights

Portable deployment: a common C++ execution path across heterogeneous robots, edge devices, and simulators.
Model-aware runtime: reusable input adapters, sequence builders, backbone execution, head plugins, and deployment adapters.
GGUF workflow: conversion and quantization tools for supported embodied models, with ready-to-use model artifacts on Hugging Face.
Evaluation-ready: integrated workflows for LIBERO and RoboTwin.

Supported Models

Model family	Current support
VLA	pi0.5 and HY-VLA
World-action model	Preliminary LingBot-VA path
Deployment backends	CPU, CUDA GPU, and heterogeneous edge hardware

The project is actively expanding model coverage and deployment backends. See the repository roadmap for the current status.

Evaluation

LIBERO: simulation workflows for manipulation tasks and smoke tests.
RoboTwin: native C++ execution for dual-arm robotic manipulation evaluation.

Detailed setup and evaluation instructions are maintained in the project repository.

Quick Start

git clone https://github.com/SEU-PAISys/Embodied.cpp.git
cd Embodied.cpp
./patches/init_third_party.sh

Then choose the model and backend you need, download a compatible GGUF checkpoint from the model collection, and follow the repository’s build instructions for CPU or CUDA execution.

Model Conversion

The repository includes conversion and quantization scripts for pi0.5, HY-VLA, and LingBot-VA. If you are adapting a new checkpoint, start from the scripts in scripts/. For standard use, the pre-converted GGUF releases are the fastest path.

Citation

@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},
  doi={10.48550/arXiv.2607.02501},
  url={https://arxiv.org/abs/2607.02501}
}

Acknowledgements

Embodied.cpp builds on and interoperates with llama.cpp, vla.cpp, OpenPI, HY-VLA, LingBot-VA, LIBERO, and RoboTwin. Each upstream project and model remains subject to its own license.

Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots