Efficient Autonomous Navigation for Air-Ground Robots in Complex Dynamic Environments
As the application scope of air-land amphibious robots (AGR) continues to expand in fields such as emergency rescue and urban inspection, the challenge of achieving autonomous navigation in complex dynamic environments has become increasingly prominent. In this regard, Wang Junming of the University of Hong Kong is based on AMOVLABP600UAVThe platform independently built a set of amphibious amphibious robots, using the Prometheus open-source framework to complete algorithm simulation verification and real-machine deployment. Following the research results of AGRNav [ICRA’24] and HE-Nav [RA-L’24], Wang Junming’s recently published papers OMEGA [RA-L 2025] and OccRWKV [ICRA 2025] provide innovative ideas for solving autonomous navigation problems in complex dynamic environments. These studies not only expand the application scenarios of air-ground amphibious robots, but also lay a theoretical foundation for improving their adaptability in uncertain environments.
Air-ground amphibious robot series content
OMEGA navigation system
Traditional air-ground robot navigation systems rely on the 3D semantic occupancy network to predict occlusion areas and perform path planning through ESDF (Euclidean signed distance field). However, in dynamic scenes (such as crowded areas), this solution will suffer from problems such as perception delay and inefficient planning.OMEGA is the first specially designed for air-ground amphibious robots to ensure efficient autonomous navigation in highly obscured and rapidly changing environments.
Technical Highlights
The OMEGA system achieves end-to-end optimization through the OccMamba perception network and the AGR-Planner planner.
OccMamba Network:
-
A three-branch network structure is innovatively proposed to decouple semantic and geometric prediction into different parts, and efficiently perform 3D semantic occupancy prediction by integrating Sem-Mamba and Geo-Mamba modules.
-
The Mamba state space model is introduced to capture long-distance dependencies using linear complexity.
-
Features are projected to BEV space fusion, and the computing load is reduced by 37% (22.1 FPS real-time inference).
AGR-Planner Planner:
-
ESDF-Free path search: Combining the Kinodynamic A* algorithm and gradient optimization, the planning time is shortened from 6.5s to 0.8s.
-
The energy consumption constraint is introduced to increase the decision-making efficiency of air-to-ground mode switching by 3 times.
Real machine verification
-
On the SemanticKITTI data set, OccMamba’s mIoU reaches 25.0 (SOTA)
-
The measured dynamic environment planning takes only 0.8 seconds, with a success rate of 98%
open-source link
https://jmwang0117.github.io/OMEGA/
OccRWKV 3D Semantic Occupancy Network
Computational redundancy is common in existing 3D semantic occupancy networks, and OccRWKV is the first 3D semantic occupancy network designed based on the RWKV architecture, which can achieve global feature modeling with linear complexity.
technological breakthrough
-
Dual branch decoupling design: Separate semantic prediction (identifying object categories) and geometric prediction (judging occupied space) into independent branches to avoid feature interference.
-
RWKV attention mechanism: Through a recurrent network structure similar to human brain memory, in a three-dimensional map containing 256×256×32 three-dimensional units, cross-regional feature association is achieved with linearly increasing calculations.
- BEV space projection: Three-dimensional features are compressed into a bird's-eye view space and fused, reducing the amount of calculation by 78.5%.
Experimental verification
-
The deployment speed is 22.2 FPS (Jetson Xavier NX) to meet the real-time needs of robots.
-
The zero-sample migration experiment shows that the movement time in the unknown environment navigation task is reduced by 16.5%.
