Odin Navigation Stack 2

ONS2 Overview

Built on the Odin1 (Odin's Eye) spatial memory module, ONS2 is an open-source, pluggable navigation stack for industrial inspection—covering mapping, localization, planning, and execution, with live 3DGS monitoring dashboard.

ONS2 system architecture: inputs and profile, perception backbone, pluggable planning stack, execution and ecosystem

Figure 1: ONS2 system architecture—unified robot profile, ego-centric perception, pluggable planners, and execution ecosystem.

Open full-size image ↗
ONS1 ONS2

What's New in ONS2

Upgrades that take the Odin Navigation Stack beyond ONS1.

01Dramatically faster planning

Plans substantially faster even on low-compute platforms, with smoother execution and nimbler obstacle avoidance.

02Upgraded VLM, bring your own model

Connect a more capable vision-language model to your own workstation for richer scene understanding.

03Built for industrial inspection

Extensive optimizations tuned to real industrial inspection scenarios and long patrol routes.

04End-to-end, fully open source

A complete inspection workflow — simulation to real robots, algorithms to the monitoring platform. See the Real2Sim tutorial.

05Plugin architecture

The stack is built around a plugin-based design, making it straightforward to integrate custom planners.

More on the way

Further capabilities will land in upcoming releases.

Odin1 On-Robot

Odin1 deployed on humanoid, quadruped and wheeled robots across real-world scenarios

Working with Odin1

Q9000 (.lx) or Odin1 (.olx) captures the site; MindCloud exports three branches—navigation maps, a live 3DGS dashboard, and Isaac Sim scenes—for ONS2 navigation, monitoring, and validation.

ManifoldTech workflow diagram: mapping, navigation and simulation

Figure 2: end-to-end workflow from site capture to ONS2 deployment.

Open full-size diagram ↗

This project showcases multiple demo pipelines built on Odin1. Each demo can be used independently or combined as an integrated stack.

01SLAM

Build 3D maps for fast, persistent relocalization, providing positioning anchors for navigation.

It may take some time to load the 3D model.

02Navigation

Long-range navigation through complex skybridges connecting multiple buildings: one complete round trip of about 640 m, with ONS 2 maintaining precise localization and stable online navigation, free of localization drift or map misalignment.

03Real2Sim Navigation

Replicate real-world sites in Isaac Sim from Q9000 or Odin1 scans, import robot models (e.g., URDF), and run navigation and manipulation tasks.

04Navigation in 3DGS Reconstructed Scenes

Built from Q9000 scanned 3DGS maps, the ONS 2 dashboard also runs in photorealistic reconstructed scenes: here a Go2 quadruped descends an outdoor staircase in simulation, with the viewpoint free to move.

3DGS scene with Go2 robot

More Exciting Features in Development

ONS2 keeps growing. Beyond this release, we are actively developing experimental modules—including multi-floor navigation and human-aware motion planning. Stay tuned for updates!

Acknowledgements

Thanks to the excellent work by ROS Navigation, NeuPAN, Ultralytics YOLO and Qwen.

Special thanks to hanruihua, KevinLADLee and bearswang for their technical support.

License

All code is released under the GNU GPL v3.0.