Source-linked AI summary
Opportunities of Self Supervised Learning for GNSS: Evaluation of a Deep Learning-Enhanced PVT Algorithm
Thomas Barbero, Bertrand Ekambi
TL;DR
Dense urban multipath causes GNSS positioning errors and is difficult to mitigate. The paper combines supervised prediction of code corrections and uncertainty with JEPA-based self-supervised pretraining, integrating outputs into a PVT estimator. The resulting DLE-PVT substantially improves accuracy, especially under harsh unseen conditions, while limited labelled diversity remains a constraint.
Problem
Dense urban reflections create multipath interference, while GNSS observable variation and costly ground-truth labeling limit robust deep-learning deployment.
Method
The method adapts JEPA to GNSS observables, then uses a heteroscedastic supervised objective to predict code corrections and uncertainty for a WLS-based PVT estimator.
Results
DLE-PVT substantially improves PVT accuracy, with SSL+FNT outperforming supervised-only training substantially under Heavy OOD conditions.
Takeaways & Limitations
The results support the potential of unlabelled GNSS data and self-supervised representations for reliable predictions under harsh, unseen urban conditions.
Takeaways & Limitations
Both labelled and unlabelled data diversity limit performance because the decoder lacks labelled examples of extreme multipath magnitudes.
Abstract
from arXiv · showhide
This work proposes a Deep Learning Enhanced PVT algorithm to mitigate multipath interference in dense urban areas. A supervised objective jointly predicts range corrections and uncertainty, while a JEPA-based self-supervised pretraining stage improves representation quality. The algorithm is evaluated over diverse driving scenarios, substantially improving PVT accuracy, particularly for unseen harsh urban conditions. These results highlight the potential of unlabelled GNSS data to improve generalization performance.
1 Introduction
Dense urban multipath creates major GNSS positioning errors and is difficult to mitigate because it depends on surroundings. The paper adapts JEPA-based self-supervised learning to GNSS observables and combines it with supervised correction and uncertainty prediction for improved generalization.
- Motivation: Building reflections create delayed GNSS replicas that produce multipath interference, a leading source of dense-urban positioning errors.Mitigation is difficult because multipath depends on the receiver’s surroundings.
- Motivation: Deep learning can exploit multipath signatures and spatial correlations, but hardware, software, and urban-environment variation complicate deployment.Larger and more diverse training datasets improve generalization, while ground-truth labeling is costly.
- Research gap: Self-supervised learning pretrains deep networks on large unlabelled receiver datasets before task-specific supervised fine-tuning.The paper identifies GNSS adaptation of extensive pretraining and masked modelling as under-explored.
- Approach: The method adapts JEPA to GNSS observables and uses a heteroscedastic objective to jointly predict code corrections and uncertainty.Predictions are integrated into a Weighted Least Squares estimator for the enhanced PVT solution.
- Evaluation: The algorithm is evaluated on two open-source datasets and 24 hours of driving data spanning diverse urban environments.The paper reports substantial accuracy gains and reliable predictions under harsh, unseen conditions.
2 Deep Learning Enhanced PVT Algorithm
The DLE-PVT algorithm extracts contextual and multipath-related GNSS features, predicts corrections and uncertainty with a Transformer, and incorporates them into a multipath-aware WLS-style estimator.
- Algorithm: DLE-PVT combines a multipath characterization module with a Multipath-Aware Positioning Algorithm, while DNN training remains separate.The positioning component uses the learned outputs during estimation.
- Feature extraction: Features include contextual measurements and multipath indicators such as CN0, range-rate consistency, Doppler residuals, and single-differenced code residuals.These indicators have established or operational correlations with multipath.
- Deep learning modelling: A Transformer processes token sequences to predict code correction and uncertainty for every token in the sliding window.Its encoder provides general representations, while the decoder extracts specialized correlations.
- Positioning: The estimator formulates rover position, velocity, clock drift, and constellation-specific time bias in a single-epoch factor graph equivalent to WLS.The formulation supports possible future extension to multi-epoch estimation.
- Positioning: Predicted corrections are applied to single-differenced code observables, while predicted uncertainties define noise models for associated factors.Excluded tokens remain in the estimator with fixed 500m standard deviation, preserving satellite geometry while reducing their contribution.
3 Deep Neural Networks Training
The paper compares supervised-only training with JEPA-style self-supervised pretraining followed by supervised fine-tuning. Both routes use a shared objective that jointly learns multipath corrections and uncertainty.
- Training routes: Supervised Only Training and SSL followed by Supervised finetuning produce identical final architectures predicting code corrections and uncertainties.The supervised objective is used alone for SPVO and during fine-tuning for SSL+FNT.
- Self-supervised learning: Masked Modelling pretraining masks a continuous azimuthal sector covering 25% of the sky and predicts masked tokens from unmasked tokens.Tokens below 40° elevation within the selected sector are masked.
- Self-supervised learning: The self-supervised loss compares predicted and target embeddings for masked tokens and adds VICReg regularization to constrain embedding distributions.The regularization addresses representation collapse in the two-branch architecture.
- Supervised learning: The supervised objective uses reference multipath approximations and heteroscedastic Laplace regression to jointly guide correction and uncertainty predictions.The reference approximation is derived from single-differenced code observables after removing geometric and clock components.
- Training routes: SSL+FNT transfers encoder weights from pretraining, reinitializes the decoder and prediction head, and trains the architecture on the supervised objective.The encoder is frozen while the remaining components are trained from scratch during fine-tuning.
4 Empirical Setting
The empirical setting combines proprietary and open-source urban driving data with scenario-level separation into in-distribution and out-of-distribution categories. Experiments isolate self-supervised pretraining and assess the effect of adding dataset diversity.
- Datasets: The study combines 24 hours of Toulouse driving data with PPC and UrbanNav datasets covering Tokyo and Hong Kong urban conditions.All datasets use mid-range u-blox receivers and multiple GNSS constellations.
- Datasets: The scenarios are grouped into Circuit #1, Circuit #2, and OS according to driving setting and dataset origin.Circuit #2 contains deep-urban scenarios with intense roadside-tree shadowing, while OS covers harsh urban conditions.
- Evaluation design: A strict scenario-level split excludes training and validation scenarios from testing to prevent data leakage.Held-out Circuit #1 scenarios are ID, Circuit #2 is Slight OOD, and the open-source datasets are Heavy OOD.
- Evaluation design: Accuracy is measured as the average of the 50th and 95th percentiles of 3D positioning error.The metric combines typical and degraded positioning performance in one score, with repeated runs using different random seeds.
- Experiments: The first experiment compares supervised-only and SSL-plus-fine-tuning models on a fixed 1M-token Circuit #1 training set.The second varies whether additional data enters pretraining, fine-tuning, or both.
5 Empirical Findings
DNN predictions consistently improve PVT accuracy over the initial WLS baseline, with SSL+FNT providing the largest benefit in harsh unseen urban conditions. Analysis links this robustness to coherent range predictions, grounded uncertainty estimates, spatially structured attention, and stage-dependent effects of training-data diversification.
- 5.1 Empirical performance: DNN predictions substantially and consistently improve PVT accuracy over the initial WLS baseline across training methods.For ID and Slight OOD scenarios, the baseline is already accurate and the two training methods perform similarly.
- 5.1 Empirical performance: SSL+FNT substantially outperforms SPRVO on Heavy OOD scenarios, increasing resilience when the initial PVT degrades severely.The harsh urban conditions cause severe baseline degradation.
- 5.1.2 Impact of Pretraining at Predictions Level: Heavy-tailed weighted range residuals for SPRVO nearly disappear with SSL+FNT, indicating more coherent predictions under harsh conditions.The SPRVO tails indicate that substantial multipath is not mitigated at the range level.
- 5.1.2 Impact of Pretraining at Predictions Level: SPRVO correction predictions correlate more strongly with initial WLS code residuals than SSL+FNT predictions, 0.96 versus 0.91.SSL+FNT uncertainty predictions correlate meaningfully with several features, whereas SPRVO uncertainty remains loosely correlated.
- 5.1.3 Attention Analysis: Attention is spatially structured: encoder layers shift from high- to low-elevation tokens, while the final decoder layer emphasizes the current token’s azimuthal sector.Low-elevation regions are associated with more intense multipath, and same-sector attention supports spatially coherent predictions.
- 5.2 Increasing Training dataset diversity: Diversification benefits depend on the training stage: pretraining helps Slight OOD most, whereas fine-tuning drives the largest Heavy OOD improvement.Pretraining diversification provides negligible Heavy OOD benefit in comparison.
- 5.2 Increasing Training dataset diversity: Diverse representations alone are insufficient when labelled data are narrow, because the decoder cannot learn nuanced predictions for extreme multipath magnitudes.The paper identifies both labelled and unlabelled data diversity as current limitations.
6 Conclusions
The DLE-PVT algorithm substantially improves positioning accuracy, notably under unseen conditions. Its supervised objective and SSL pretraining contribute complementary benefits, with SSL improving prediction coherence through higher-quality representations and robust features.
- DLE-PVT substantially improves positioning accuracy, notably under unseen conditions.
- The supervised objective predicts code corrections and uncertainty for multipath-aware positioning.
- SSL pretraining improves prediction coherence through higher-quality representations and reliance on robust features.
- Future work will examine whether scaling unlabelled data alone can drive further generalization gains.