Current State (Academic Reproduction):
The code currently available in the icra2026 branch contains the experimental implementation corresponding to our ICRA 2026 paper. It is provided for algorithm verification and academic reference. It is not yet optimized for direct production deployment.
Upcoming Release (June 2026):
We are currently conducting a comprehensive system refactoring based on FineNav-Engine—a dedicated C++20 development framework for robotics navigation. This upcoming release will be fully optimized for direct production deployment and accompanied by detailed documentation.
If you are interested in our work, please ⭐Star this repository to receive updates on the upcoming full release ;D
FineNav is a navigation framework tailored for ground robots operating in unstructured 3D environments. It feaures on a cache-memory-like hierarchical mapping system, which achieves a favorable balance between low-latency real-time perception (for dynamic obstacle avoidance and terrain analysis), and scalable global storage (for large-scale 3D reasoning).
Video Demonstration: Bilibili | Youtube
The FineNav navigation framework is capable of dealing with various scenarios in unstructured environments within a unified pipeline, including:
The core mapping system leverages a hierarchical architecture to balance latency and storage:
-
Ring-buffer based Local Grid: Achieves
$O(1)$ spatial shifts, enabling high-rate map updates and low-latency reactive perception. - Global OctoMap: Maintains a persistent, memory-efficient 3D representation for large-scale spatial reasoning and cross-floor path planning.
The interaction mechanism between both maps is specifically designed to mirror a cache-memory model, effectively decoupling the real-time perception pipeline from costly global map updates:
FineNav is built upon a highly modular architecture of reusable components, ensuring that each module can be replaced or reconfigured independently. Specifically, terrain analysis is tightly coupled with locomotion capability and is therefore exposed as a plugin. Developers can dynamically load custom analyzers without modifying the core system.
The code currently available in the icra2026 branch contains the experimental implementation corresponding to our ICRA 2026 paper. It is provided for algorithm verification and academic reference. It is not yet optimized for direct production deployment.
The production-ready version based on FineNav-Engine will be released soon.
If you find this work helpful in your research, please cite our paper:
@inproceedings{wang2026finenav,
title={FINENAV: A Versatile Framework Enhancing Ground Robot Navigation in Unstructured Environment},
author={Wang, Jinghui and Wang, Chenyang and Cao, Yuxuan and Sun, Zelong and Xi, Wang and He, Jianping},
booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
year={2026}
}