"coVoxSLAM: GPU Accelerated Globally Consistent Dense SLAM"

Full GPU acceleration makes dense SLAM (Simultaneous localization and mapping) 140x faster while maintaining accuracy

SLAM (Simultaneous Localization and Mapping) is a technology that enables robots to create maps of unknown environments while simultaneously tracking their own position within that map.

📚 https://arxiv.org/abs/2410.21149

🎯 Original Problem:

Dense SLAM (Simultaneous localization and mapping) systems for mobile robots face computational challenges in real-time processing and global map consistency, especially in large-scale environments. Current systems struggle with GPU utilization and memory efficiency.

-----

🔧 Solution in this Paper:

• Introduces coVoxSLAM - a fully GPU-accelerated volumetric SLAM system

• Uses raycasting instead of projection mapping for point cloud integration

• Implements Structure of Arrays (SoA) data layout for better memory efficiency

• Features a multi-step approach for point cloud integration with balanced GPU thread workload

• Employs GPU-optimized pose graph optimization for backend processing

• Uses hash tables with efficient collision handling for spatial data indexing

-----

💡 Key Insights:

• Raycasting outperforms projection mapping for TSDF integration

• SoA data layout significantly improves memory throughput

• GPU-based backend processing can match CPU accuracy while being faster

• Hash table performance remains stable until 85% loading factor

-----

📊 Results:

• 30x to 140x faster TSDF integration compared to Voxblox

• 10x to 50x faster ESDF computation

• 2x to 14x faster backend processing

• 1.5x to 2x speedup compared to nvBlox

• Maintains RMSE accuracy between 0.8-1.2 meters