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

Abstract: Deploying embodied AI models remains fragmented across model-specific Python stacks, backend assumptions, and robot-side code. Embodied.cpp is a portable C++ runtime that captures a shared execution path for vision-language-action and world-action models through input adapters, sequence builders, backbone execution, head plugins, and deployment adapters. It supports multi-rate execution, latency-first fused inference, and extensible operators and interfaces across heterogeneous devices, robots, and simulators. Across three VLA and two WAM models, it achieves 1.05–2.70x speedups and 7–77% lower VRAM than Python baselines while preserving control quality.

@article{embodiedcpp,
  title={Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots},
  author={Ling Xu and Borui Li and Hao Wu and Chuyu Han and Xiangyu Li and Mohan Hua and Shiqi Jiang and Ting Cao and Chuanyou Li and Sheng Zhong and Shuai Wang},
  journal={arXiv preprint arXiv:2607.02501},
  eprint={2607.02501},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2607.02501},

  year={2026}
}