PaperlayerView exact revision on arXiv

Source-linked AI summary

Transformer-based Monte Carlo Localization in Construction Meshes

Linus Kramer, William Talbot, Olga Vysotska, Marco Hutter

arXiv:2609.31357v1cs.RO

TL;DR

Construction-site localization is challenged by unavailable indoor GPS and self-similar places. The paper trains a LiDAR place-recognition model on synthetic mesh scans and uses its uncertainty-aware predictions in MCL. On real-world datasets, it outperforms diffusion and ScanContext++ baselines across datasets.

  • Problem

    Construction-site global localization is difficult because indoor GPS is often unavailable and self-similar or feature-poor places challenge disambiguation from a single observation.

    Source: §I · p. 1

  • Method

    The system trains a LiDAR place-recognition model on synthetic mesh scans and uses its uncertainty-aware predictions as observation weights in MCL.

    Source: §I · p. 2, §III · p. 3, E. Monte Carlo Localization · p. 5

  • Results

    The confidence-prediction decoder improved 1 m recall over diffusion and ScanContext++ across datasets; on Hilti it reached 91% versus 43% and 29%, respectively.

    Source: D. Monte Carlo Localization Evaluation · p. 7

  • Takeaways & Limitations

    The system generalizes from synthetic training to real scans in the same environment and is reported to outperform diffusion-based and ScanContext++ baselines without real-world training data.

    Source: §V · p. 8

    The system depends on manually acquired reality-captured meshes rather than BIM models, requiring an additional data-collection step before deployment.

    Source: §V · p. 8

Abstract

from arXiv · show

To be able to perform inspection or digitization tasks, mobile robots on construction sites must be able to localize themselves reliably with respect to a global reference frame that is shared with a building map. Similar room layouts and low-texture surfaces pose a challenge for existing LiDAR- and vision-based localization methods. We approach this problem with a LiDAR-based global relocalization system that estimates the robot's pose relative to a building mesh and combines a PointNet++ encoder with a place recognition decoder, whose outputs serve as a learned observation model within a Monte Carlo Localization (MCL) framework. The pipeline is trained exclusively on synthetic LiDAR scans obtained by simulating the robot's sensors inside the building mesh. Our approach is robust in ambiguous environments due to an uncertainty-aware decoder that scales positional likelihoods and a resampling strategy that injects model hypotheses into the particle set, enabling recovery from potential particle depletion. Evaluations on real-world datasets show that our method outperforms both diffusion-based and ScanContext++ baselines while maintaining fast inference (18 ms per call), demonstrating the practicality of synthetic-data training for mesh-referenced global localization in construction robotics.

I. INTRODUCTION

Automating construction-site inspection depends on reliable global localization, but similar, feature-poor environments challenge single-observation methods. The paper combines learned LiDAR place recognition with MCL to track location hypotheses over time.

I. INTRODUCTION

  • Global localization lets a robot start an inspection without a known initial location and recover its map position if tracking fails.Source: §I · p. 1
  • Construction sites challenge localization through unavailable indoor GPS, self-similar places, sparse helpful features, and changing differences between maps and observations.Source: §I · p. 1
  • MCL tracks multiple pose hypotheses over time, addressing ambiguity that makes a single LiDAR observation difficult to disambiguate.Source: §I · p. 1
  • The system combines a trained place-recognition model predicting locations from one LiDAR scan with a particle filter using observations and temporal state; it trains on synthetic scans and generalizes to real scans from the same environment.Source: §I · p. 2
  • The contributions include scan-based uncertainty in the MCL observation model and a resampling strategy for recovery when the filter becomes overconfident in an incorrect location.Source: §I · p. 2

II. RELATED WORK

Prior localization approaches include scan descriptors, learned embeddings, overlap estimation, pose regression, diffusion, and ICP-based scan-to-map methods. This work combines learned place recognition with MCL to track multiple hypotheses in construction meshes.

II. RELATED WORK

  • ScanContext++ uses polar-coordinate bins for scan descriptors, while MinkLoc3D and PointLoc use learned point-cloud representations; PointLoc predicts one pose without hypothesis tracking.Source: §II · p. 2
  • OverlapNet estimates scan-pair overlap for MCL, but requires a simulated scan for every particle at every update, making it computationally expensive and less suited to real-time deployment.Source: §II · p. 2
  • The diffusion approach iteratively refines an initial position and can produce multiple hypotheses; the paper adopts and extends its mesh-based LiDAR synthesis pipeline and uses it as a baseline.Source: §II · p. 2
  • MCL can become stuck after particles converge on an incorrect hypothesis; the proposed approach injects learned place-recognition locations to support recovery.Source: §II · p. 2
  • Construction-site ICP approaches produce a single pose per observation, whereas the proposed learned observation model and particle filter track multiple hypotheses over time.Source: §II · p. 2

III. LOCALIZATION IN CONSTRUCTION SITE MESHES

  • The pipeline trains on simulated mesh LiDAR scans, then uses place-recognition predictions as uncertainty-aware observation weights within MCL alongside motion, resampling, and prediction.Source: §III · p. 3, Figure 2 · p. 3

A. Robot Simulation

The training pipeline simulates LiDAR observations from realistic robot poses in the mesh, then encodes each point cloud into a global geometric descriptor.

A. Robot Simulation

  • Valid robot poses are sampled from standable mesh surfaces using a 0.1 m horizontal grid and ceiling-height checks, accommodating stairs and ramps that simple plane detectors miss.Source: A. Robot Simulation · p. 3
  • For each pose, ray tracing generates a LiDAR cloud with noise in ray directions and distances; points outside 0.5–20 m are removed and 4096 points are randomly retained.Source: A. Robot Simulation · p. 3

B. Point Cloud Encoder

  • A PointNet++ encoder groups points into progressively larger regions, learns local geometry, and max-pools the features into a global descriptor.Source: B. Point Cloud Encoder · p. 3

C. Place Recognition Decoder

The place-recognition decoder predicts a distribution over robot positions rather than a single estimate, allowing it to represent multiple plausible locations from an ambiguous scan.

C. Place Recognition Decoder

  • The decoder predicts a probability density over continuous 3D position, using five weighted Gaussian components to represent alternative location hypotheses.Source: C. Place Recognition Decoder · p. 3
  • Five trainable hypothesis queries act as slots for alternative locations, and the decoder uses four transformer layers with 256-dimensional features and four attention heads.Source: C. Place Recognition Decoder · pp. 3–4
  • Multiple hypotheses defer a fragile single-location choice until future observations can resolve ambiguity in self-similar rooms or featureless corridors.Source: C. Place Recognition Decoder · p. 4
  • Linear heads predict component means, standard deviations, and softmax mixture weights; training minimizes the negative log-likelihood of the ground-truth position under the mixture.Source: C. Place Recognition Decoder · p. 4
  • Computing the training likelihood in log-space prevents underflow when the ground-truth position has very small likelihood.Source: C. Place Recognition Decoder · p. 4

D. Place Recognition Decoder with Confidence Prediction

The decoder adds a confidence prediction that adjusts how much its place-recognition likelihood influences localization. It can rely more on a uniform likelihood in ambiguous regions and more on its prediction in distinctive areas.

D. Place Recognition Decoder with Confidence Prediction

  • The confidence head predicts β ∈[0, 1] to express confidence in the decoder’s mixture-of-Gaussians prediction.When β is near 1, the mixture is trusted; when β is near 0, the model falls back to a uniform likelihood over valid mesh positions.Source: D. Place Recognition Decoder with… · p. 4
  • The uniform fallback assigns equal likelihood to every valid mesh position, approximated using occupied voxels at 1 m resolution.Its likelihood is the inverse of navigable mesh volume V.Source: D. Place Recognition Decoder with… · p. 4
  • Training encourages β to favor whichever likelihood—the mixture prediction or uniform fallback—better explains the ground truth for the scan.The model can reduce β in ambiguous, self-similar regions and increase it in distinctive areas.Source: D. Place Recognition Decoder with… · p. 4

E. Monte Carlo Localization

The MCL filter tracks 1,000 feasible pose hypotheses using motion and learned observation weights. Its confidence scaling and hypothesis-injecting resampling are designed to preserve alternatives and recover from mode collapse.

E. Monte Carlo Localization

  • The filter represents pose belief with 1,000 particles initialized at physically feasible poses inside the mesh.Each particle encodes a 6-DoF pose: 3D position and quaternion orientation.Source: E. Monte Carlo Localization · p. 4
  • For human-recorded trajectories, motion is approximated from ground-truth pose changes with zero-mean Gaussian noise scaled to motion and elapsed time.A legged-inertial estimator can provide smooth odometry when available.Source: E. Monte Carlo Localization · p. 4
  • The learned decoder supplies particle weights by combining its predicted position likelihood with a uniform prior, scaled by scan confidence.The sim-to-real baseline confidence is 0.5; learned β scales that baseline, and lower confidence reduces weight variation so resampling retains more hypotheses.Source: E. Monte Carlo Localization · p. 5
  • Systematic resampling draws particles from normalized weights using a random start and fixed intervals across their cumulative weight stack.This common particle-filter method can retain some low-probability particles.Source: E. Monte Carlo Localization · p. 5
  • After resampling, five particle positions are replaced with Gaussian means predicted by the place-recognition model to keep alternative modes represented and enable escape from mode collapse.Without this intervention, a filter concentrated on one hypothesis may recover only through motion-model randomness, which is often too weak.Source: E. Monte Carlo Localization · p. 5
  • DBSCAN selects the largest cluster’s centroid as the predicted position, falling back to the mean of all particles if no valid cluster is found.Clusters use a 1 m neighborhood radius and a minimum size of 100 particles.Source: E. Monte Carlo Localization · p. 5

IV. EXPERIMENTAL EVALUATION

The evaluation supports claims that the MCL approach outperforms baselines in construction environments, confidence estimation improves robustness, and the resampling strategy enables recovery from an incorrect mode.

IV. EXPERIMENTAL EVALUATION

  • The MCL-based approach consistently outperforms the baselines in construction environments.The evaluation also reports that confidence estimation improves robustness under self-similarity, clutter, and incomplete structures.Source: §IV · p. 5
  • Confidence estimation improves robustness under self-similarity, clutter, and incomplete structures that make construction-site localization difficult.Source: §IV · p. 5
  • The proposed resampling strategy enables recovery when the particle filter collapses into an incorrect mode.Source: §IV · p. 5

A. Datasets

The evaluation uses four real-world datasets spanning different scales, floor counts, and acquisition setups; three are from active construction sites, and one provides a non-site test of strong self-similarity.

A. Datasets

  • Figure 4 marks the staircase used in multifloor datasets in green and includes a 10 m scale bar; the Aesch view includes a representative environment image.Source: Figure 4 · p. 5
  • The four datasets span small to large indoor volumes, one to three floors, and robot or handheld trajectories; meshes were captured on-site and used as maps.Ground-truth trajectories were estimated with Open3D SLAM using preexisting maps, with an accurate initial pose and offline tuning, so this was not a viable online method.Source: A. Datasets · p. 5
  • Hilti’s exp06 sequence covers a mostly open indoor area with few partitions and limited structural cues, serving as a low-self-similarity baseline.The passage reports a volume of 1193 m3 and Hesai PandarXT-32 trajectory LiDAR.Source: A. Datasets · pp. 5–6
  • The Apartment dataset covers two floors of an active construction site, with smaller rooms and greater structural complexity than Hilti.Its volume is 1184 m3; the trajectory used an ANYmal robot and the map was captured with a handheld Leica BLK2GO.Source: A. Datasets · p. 6
  • The Stairwell dataset is the only non-construction-site environment and tests strong self-similarity across three floors linked by a staircase around an elevator.It has bare concrete walls without furniture and a volume of 2470 m3.Source: A. Datasets · p. 6
  • Aesch is the largest and most self-similar setting, with three floors and near-identical room layouts on two floors.Its active construction site spans 9719 m3, providing a test of scalability and robustness under strong ambiguity.Source: A. Datasets · p. 6

B. Baseline

The MCL evaluation reports higher 1 m recall for the confidence-enabled transformer decoder than for diffusion and ScanContext++ across four datasets, with a similar pattern under regular particle reinitialization. The section also describes the baselines, training setup, confidence ablation, and particle reseeding used to recover from mode collapse.

B. Baseline

  • The baselines are ScanContext++ with a simulated descriptor database, diffusion with five hypotheses, and a random predictor sampled from the simulated training set.For MCL, the top five ScanContext++ matches and five diffusion hypotheses are converted to MoGs using the same fixed parameters.Source: B. Baseline · p. 6

C. Implementation Details

  • Training uses Adam at a 1 × 10−4 learning rate and batch size 16, with dataset-specific schedules of 400–1,000 epochs on one NVIDIA TITAN RTX GPU.The plain decoder has 2.9M trainable parameters and an estimated model size of 11.5 MB.Source: C. Implementation Details · p. 6

D. Monte Carlo Localization Evaluation

  • 1 m recall is 91% on Hilti, 86% on Apartment, 73% on Stairwell, and 42% on Aesch, exceeding diffusion (43%, 69%, 51%, 14%) and ScanContext++ (29%, 0%, 13%, 0%).The improvement follows the same trend in Reinit, where particles restart every 10 steps, so it does not depend on favorable initial conditions.Source: D. Monte Carlo Localization Evaluation · p. 7
  • Removing confidence prediction lowers 1 m recall by 42, 5, 12, and 9 percentage points on Hilti, Apartment, Stairwell, and Aesch, respectively.Similar relative differences occur in Reinit, supporting a consistent benefit from data-driven confidence prediction in the MCL pipeline.Source: D. Monte Carlo Localization Evaluation · p. 7
  • Resampling injects predicted place hypotheses at the means of the MoG, repopulating alternative modes as observations arrive and shifting probability mass toward the correct region.Standard MCL relies only on motion-model noise to escape an incorrect mode collapse.Source: D. Monte Carlo Localization Evaluation · p. 7

E. Place Recognition Evaluation

The place-recognition evaluation finds the confidence-enabled transformer decoder has the highest recall at all reported radii. Its confidence scores align with localization error, and inference is faster than the diffusion baseline.

E. Place Recognition Evaluation

  • 0.018 s per call is an order of magnitude faster than the diffusion baseline’s 0.11 s per call.The authors cite this inference time as supporting suitability for real-time deployment.Source: E. Place Recognition Evaluation · p. 8
  • The confidence-enabled transformer decoder achieves the highest recall at 2 m, 1 m, and 0.5 m in place recognition.The evaluation compares predicted distributions through their maximum a posteriori estimate; the text attributes the improvement over the decoder without confidence prediction likely to the confidence-augmented loss.Source: E. Place Recognition Evaluation · pp. 7–8
  • Lower predicted β values occur in ambiguous hallways and staircases, coinciding with higher map-estimate errors.The confidence visualization pairs place-prediction error colored from 0 m to 4 m with β values from 0 to 1.Source: E. Place Recognition Evaluation · p. 8, Figure 6 · p. 8

V. CONCLUSION

The conclusion presents a transformer-decoder place-recognition model combined with adapted MCL for global localization in construction sites. It reports superior performance to diffusion-based and ScanContext++ baselines while identifying mesh acquisition and rapidly changing environments as limitations.

V. CONCLUSION

  • The localization pipeline combines a transformer-decoder place-recognition model with adapted MCL and requires a reality-captured building mesh rather than real-world training data.The system is designed for self-similar construction environments and recovery from mode collapse.Source: §V · p. 8
  • Manually acquiring reality-captured meshes adds a data-collection step before deployment, and the method has not been thoroughly evaluated in rapidly changing environments with possible mesh–LiDAR discrepancies.The authors contrast these meshes with BIM models, noting the captured meshes provide an accurate reference and simplify localization.Source: §V · p. 8
Loading 2609.31357v1…