HV-LIOM: Adaptive Hash-Voxel LiDAR-Inertial SLAM with Multi-Resolution Relocalization and Reinforcement Learning for...

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Real-time 3D mapping in dynamic, GNSS-denied environments is crucial for autonomous systems.
• Existing LiDAR-inertial SLAM methods face challenges with memory efficiency, robustness to initialization errors, and efficient exploration.

Purpose of the Study:

• To present HV-LIOM, a unified framework for LiDAR-inertial SLAM and autonomous exploration.
• To improve memory efficiency, real-time state estimation, and localization robustness.
• To enhance autonomous exploration efficiency in unknown environments.

Main Methods:

• Adaptive hash-voxel mapping scheme for optimized memory usage and geometric complexity.
• Multi-resolution relocalization strategy for robust localization under large initial pose errors.
• Learning-based loop-closure and global pose-graph optimization for map consistency.
• Soft Actor-Critic (SAC) policy for online selection of informative navigation targets.

Main Results:

• HV-LIOM demonstrates improved absolute pose accuracy: up to 15.2% over FAST-LIO2 indoors and 7.6% outdoors.
• The learned exploration policy achieves comparable or superior area coverage with reduced travel distance and time.
• The framework shows enhanced performance on public datasets (Hilti, NCLT) and a custom mobile robot.

Conclusions:

• HV-LIOM provides an efficient and robust solution for real-time 3D mapping and autonomous exploration.
• The adaptive mapping and learning-based components significantly improve performance in challenging scenarios.
• The framework advances the capabilities of autonomous systems in GNSS-denied environments.