Uni-LaViRA: Language-Vision-Robot Actions Translation for Unified Embodied Navigation

Uni-LaViRA is a training-free framework for unified embodied navigation: instead of fine-tuning on millions of robot data and thousands of GPU hours, it casts navigation as a three-layer translation — language → vision → robot action — where pretrained multimodal LLMs plan in language, ground in vision, and map the result to each robot's low-level controls. One architecture spans four navigation tasks and four heterogeneous robots, yet matches or even surpasses trained SOTA navigation foundation models.

Release TODO List

• Conda environment setup
• Docker environment setup
• Simulation evaluation code
  • VLN — R2R-CE, RxR-CE
  • ObjectNav — HM3D-v2, HM3D-OVON
  • EQA — MP3D-EQA
  • Aerial VLN — OpenUAV
• Real-world deployment code
  • Agilex Cobot Magic
  • Unitree Go1
  • Unitree G1
  • Self-built UAV

Repository Layout

Path	Description
sim-code/habitat/	Habitat-based simulation for VLN + ObjectNav + EQA.
sim-code/airsim/	AirSim / TravelUAV-based simulation for aerial VLN.
real-world-code/cobot_magic/	Agilex Cobot Magic deployment.
real-world-code/unitree_g1/	Unitree G1 humanoid deployment.
real-world-code/unitree_go1/	Unitree Go1 quadruped deployment.
real-world-code/self_built_uav/	Self-built UAV (ROS) deployment.

Task × Platform

Task	Simulation	Real Robots
VLN	sim-code/habitat	cobot_magic · unitree_g1 · unitree_go1 · self_built_uav
ObjectNav	sim-code/habitat	cobot_magic · unitree_g1 · unitree_go1 · self_built_uav
EQA	sim-code/habitat	cobot_magic · unitree_g1 · unitree_go1 · self_built_uav
Aerial VLN	sim-code/airsim	self_built_uav

Each ground robot runs all of vln / object_nav / eqa / interact from one shared pipeline and a single LA/VA model; only the instruction differs.

Quickstart

• Run simulation evaluation → sim-code/habitat/README.md (indoor) or sim-code/airsim/README.md (aerial).
• Deploy on a real robot → the README inside the matching real-world-code/<robot>/ folder.

Each subfolder is self-contained: environment setup, datasets, checkpoints, and run commands all live in its own README. The simulation and real-world trees use separate environments and do not share a single install.

Citation

@article{ding2026unilavira,
  title={Uni-LaViRA: Language-Vision-Robot Actions Translation for Unified Embodied Navigation},
  author={Ding, Hongyu and others},
  journal={arXiv preprint arXiv:2605.27582},
  year={2026}
}
@article{ding2025lavira,
  title={LaViRA: Language-Vision-Robot Actions Translation for Zero-Shot Vision Language Navigation in Continuous Environments},
  author={Ding, Hongyu and Xu, Ziming and Fang, Yudong and Wu, You and Chen, Zixuan and Shi, Jieqi and Huo, Jing and Zhang, Yifan and Gao, Yang},
  journal={arXiv preprint arXiv:2510.19655},
  year={2025}
}

Acknowledgements

This work builds on LaViRA, CA-Nav, NavDP, and Depth-Anything V2. Thanks for their great work!

License

The open-sourced code is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0) — see LICENSE. Third-party submodules and vendored components remain under their own upstream licenses.