LAPS framework boosts incremental LiDAR mapping with 4.66% recall gain
New replay management method beats catastrophic forgetting in neural distance fields
Neural distance fields are compact continuous 3D representations ideal for incremental LiDAR mapping, but online optimization suffers catastrophic forgetting—new observations degrade earlier geometry. Replay-based training helps, but existing methods use passive buffers and uniform sampling, wasting memory on redundant data and undertraining poorly constrained regions.
LAPS solves this with a two-pronged approach: reliability-based active pooling selectively retains the most informative historical samples under bounded memory, while uncertainty-guided active sampling prioritizes optimization in areas with high reconstruction uncertainty. On the challenging Oxford Spires dataset (Blenheim Palace 05 sequence), LAPS improves recall by 4.66 percentage points and F1 score by 3.79 percentage points over PIN-SLAM, all while maintaining competitive geometric accuracy. The team has released an open-source implementation on GitHub. This work was accepted at RA-L 2026 and promises more robust 3D mapping for autonomous systems.
- Reliability-based active pooling retains only the most informative historical samples under limited memory constraints
- Uncertainty-guided active sampling focuses training on poorly constrained regions to reduce forgetting
- Improves recall by 4.66 pp and F1-score by 3.79 pp over PIN-SLAM on Oxford Spires dataset
Why It Matters
More robust incremental 3D mapping could improve autonomous navigation, drone surveying, and AR localization.