💻 computer science

LAPS: Improving Incremental LiDAR Mapping using Active Pooling and Sampling for Neural Distance Fields

The paper proposes LAPS, a replay management framework that enhances incremental LiDAR mapping using neural distance fields by combining reliability-based active pooling and uncertainty-guided active sampling to mitigate catastrophic forgetting and improve reconstruction completeness.

Original authors: Dongjae Lee, Wooseong Yang, Yifu Tao, Maurice Fallon, Ayoung Kim

Published 2026-05-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Dongjae Lee, Wooseong Yang, Yifu Tao, Maurice Fallon, Ayoung Kim

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a robot trying to build a 3D map of a building as it walks through it, using a laser scanner (LiDAR) to "see" the walls and furniture. The robot uses a smart, AI-based system to store this map. However, there's a major problem: as the robot learns about new rooms, it tends to "forget" the details of the rooms it saw earlier. This is called catastrophic forgetting. It's like trying to learn a new language every day but losing your vocabulary from yesterday.

To fix this, the robot keeps a "replay buffer"—a digital notebook where it stores old laser scans to review while learning new ones. But the paper argues that most robots are bad at managing this notebook. They either fill it up with too much junk (redundant data) or they spend too much time reviewing the parts of the map they already know perfectly, ignoring the blurry, uncertain spots.

The authors propose a new system called LAPS (Active Pooling and Sampling) to fix this notebook management. They use two clever strategies, which can be understood through these analogies:

1. The "Quality Control" Filter (Reliability-based Active Pooling)

Imagine the robot is taking photos of a room to put in a scrapbook.

The Old Way: The robot just takes every single photo it sees. If it stands in one spot and takes 1,000 photos of the same wall, the scrapbook gets heavy and full of duplicates, while the other side of the room (which it only glanced at once) gets no photos at all.
The LAPS Way: The robot acts like a strict editor. It looks at the photos and asks, "Is this picture clear and reliable?"
- If the robot is looking at a wall from a weird angle or far away, the photo is "noisy" and unreliable.
- If the robot is looking straight on and close up, the photo is "reliable."
- LAPS keeps only the best, most reliable photos for each tiny section of the room and throws away the blurry or redundant ones. This keeps the scrapbook small but ensures it has high-quality data for every part of the map, even the parts the robot didn't visit often.

2. The "Study Guide" Strategy (Uncertainty-guided Active Sampling)

Now, imagine the robot is studying for a test using that scrapbook.

The Old Way: The robot picks pages from the scrapbook completely at random. It might spend 10 minutes studying the front door because it's easy and it already knows it perfectly, while ignoring the confusing hallway it hasn't figured out yet.
The LAPS Way: The robot looks at its "confidence meter." It asks, "Where am I most unsure?"
- If a part of the map is blurry or missing (high uncertainty), LAPS flags it as a "priority study zone."
- The robot then focuses its learning energy specifically on those confusing areas, while still glancing at the easy parts to make sure it doesn't forget them.
- This is like a student focusing their study time on the math problems they got wrong, rather than re-reading the ones they already aced.

The Results

The paper tested this system on both computer simulations and real-world datasets (like walking through Oxford University buildings).

Completeness: LAPS built more complete maps. It filled in the "holes" and missing pieces that other systems left behind.
Accuracy: It didn't sacrifice accuracy for completeness; the maps were still geometrically precise.
Efficiency: It did all this without needing more computer memory or time than the existing top methods.

In short, LAPS makes the robot's memory smarter. It stops wasting space on duplicate photos and stops wasting time studying what it already knows, resulting in a much more complete and accurate 3D map of the world.

Technical Summary: LAPS – Improving Incremental LiDAR Mapping using Active Pooling and Sampling for Neural Distance Fields

Problem Statement
Incremental 3D mapping using neural distance fields offers a compact and continuous representation of geometry, making it attractive for large-scale robotics applications. However, online optimization of these networks is susceptible to "catastrophic forgetting," where new observations overwrite previously learned geometry, degrading reconstruction quality. While replay-based training (storing historical samples to interleave with current data) is a standard solution, existing methods rely on passive replay buffers and uniform sampling. These approaches suffer from two main limitations under fixed memory and optimization budgets:

Inefficient Retention: Passive strategies (e.g., local windows or fixed-capacity queues) often over-represent frequently scanned regions while discarding supervision from sparsely observed areas, leading to spatial sample imbalance.
Inefficient Allocation: Standard random sampling during optimization frequently revisits well-constrained regions, leaving under-observed or poorly reconstructed areas insufficiently optimized. This results in locally accurate but globally incomplete maps with missing structures.

Methodology
The authors propose LAPS (Learning with Active Pooling and Sampling), a unified replay management framework designed to optimize both the retention of historical data and the allocation of training effort. The system operates on a neural distance field parameterized by a multi-resolution sparse voxel grid.

Reliability-based Active Pooling (Retention):
To manage memory constraints and reduce spatial imbalance, LAPS filters incoming training samples before they enter the replay buffer.
- Local Windowing: Samples are first restricted to a local window around the current sensor origin.
- Voxel-wise Capping: Samples are assigned to voxels at the lowest grid resolution. Each voxel is capped at a maximum capacity ( $\tau$ ).
- Reliability Ranking: When a voxel exceeds capacity, the system retains only the $\tau$ most "reliable" samples. Reliability is estimated by calculating the expected squared error of the projective signed distance. This error model accounts for systematic bias (dependent on the incidence angle between the ray and surface normal) and range-dependent measurement uncertainty. Samples with lower expected error are prioritized, ensuring that redundant or noisy observations in dense regions do not crowd out informative samples from sparse regions.
Uncertainty-guided Active Sampling (Allocation):
To optimize the limited online training budget, LAPS directs optimization efforts toward regions where the model is most uncertain.
- Uncertainty Estimation: The system approximates model (epistemic) uncertainty using a Laplace approximation around the maximum a posteriori (MAP) estimate. It utilizes a perturbation field parameterized by a neural network and approximates the posterior covariance via the Fisher information matrix, assuming a zero-mean Gaussian prior.
- Batch Construction: Voxels in the replay buffer are classified into "uncertain" ( $\sigma_e \ge \lambda$ ) and "certain" sets. During training iterations, mini-batches are constructed by drawing a specific number of samples from the uncertain set and the remainder from the certain set. This ensures that under-constrained regions receive more training updates without completely neglecting the rest of the map.

Key Contributions

Formulation: The paper frames replay management as a critical, under-explored bottleneck in incremental neural LiDAR mapping, specifically addressing the dual constraints of fixed memory and limited online optimization budgets.
Framework: The introduction of LAPS, which integrates reliability-based active pooling to maintain a compact, balanced replay buffer and uncertainty-guided active sampling to prioritize training on under-constrained regions.
Performance: Demonstration that LAPS improves reconstruction completeness (recall) and global consistency while maintaining competitive geometric accuracy, outperforming state-of-the-art baselines on both synthetic and real-world datasets.

Experimental Results
The authors evaluated LAPS on three benchmarks: MaiCity (synthetic), Newer College, and Oxford Spires (real-world).

Quantitative Performance: On the Oxford Spires "Blenheim Palace 05" sequence, LAPS improved recall by 4.66 percentage points and the F1-score by 3.79 percentage points over PIN-SLAM, a strong neural baseline. Across datasets, LAPS consistently achieved the highest F1-scores and lower Chamfer-L1 distances compared to baselines like SHINE-Mapping, N3-Mapping, and PIN-SLAM.
Qualitative Improvements: Visual comparisons show that LAPS produces smoother, more complete surfaces with fewer "floating artifacts" and missing geometries in complex or under-observed areas compared to baselines.
Ablation Studies: Experiments confirmed that both components contribute to performance. Active pooling significantly reduced replay memory usage (by ~56% in one test) and improved completeness by mitigating spatial imbalance. Active sampling improved training efficiency, achieving higher reconstruction quality in fewer optimization iterations.
Runtime: LAPS maintains practical runtime for online mapping (approx. 0.10–0.13 seconds per frame), remaining comparable to PIN-SLAM and significantly faster than SHINE-Mapping and N3-Mapping.

Significance and Claims
The paper claims that LAPS addresses a fundamental limitation in incremental neural mapping: the inefficiency of how historical data is stored and utilized. By shifting from passive, uniform strategies to active, reliability- and uncertainty-driven management, the method achieves a better balance between memory efficiency and reconstruction fidelity. The authors position LAPS not as a new neural architecture, but as a critical management layer that enhances existing neural distance field frameworks, enabling more robust and complete 3D reconstructions in large-scale, dynamic environments. The work is presented as an open-source contribution to facilitate further research in incremental learning for robotics.

1. The "Quality Control" Filter (Reliability-based Active Pooling)

2. The "Study Guide" Strategy (Uncertainty-guided Active Sampling)

The Results

More like this