Simultaneous localization and mapping (SLAM) is a core challenge for autonomous vehicles and robots. Geometric LiDAR-inertial SLAM systems are popular because they estimate state efficiently, but they fall short when it comes to representing the environment as a continuous map or refining results globally. Neural SLAM methods offer those capabilities, yet often at a computational cost that makes real-time onboard use difficult.
The new paper introduces LIN-SLAM, a lightweight LiDAR-inertial neural SLAM system aimed at real-time onboard navigation. By combining the strengths of geometric estimation with neural map representations, the authors propose a way to get both efficiency and the richer modeling that neural approaches provide. The abstract does not give implementation details, but the stated goal is to close the gap between the two families of methods.
If successful, this could make neural SLAM practical for drones, robots, and other platforms with limited compute. The work is presented as a step toward continuous, globally consistent mapping without sacrificing the speed needed for live operation.