Abstract
Autonomous robots operating in GPS-denied environments, such as the deep ocean or dense forests, rely on dead reckoning, which inevitably suffers from cumulative drift. To address this, we propose a novel unsupervised framework for real-time geo-localization that matches high-resolution camera observations to low-resolution geospatial maps. Our approach utilizes a dual-encoder architecture trained via contrastive learning to bridge significant scale gaps (up to 70x) and cross-modality differences (e.g., RGB to SAR). By transforming maps into probabilistic "belief maps," the system enables drift-free navigation and real-time inference on edge hardware. Experimental results on underwater and aerial datasets demonstrate substantial improvements in localization accuracy and computational efficiency over existing methods.
Introduction: The Challenge of GPS-Denied Navigation
Robots in environments without GPS access typically use Inertial Measurement Units (IMUs) and cameras to track incremental movement. However, small errors in these sensors accumulate over time, leading to "drift" that can cause mission failure. Geo-localization—matching current visual observations to a pre-existing global map—is the standard solution to reset this drift.
The primary hurdle in geo-localization is the "cross-scale gap." A robot's onboard camera captures fine-grained, high-resolution details, whereas available geospatial maps (satellite imagery, bathymetric charts) are often low-resolution. Furthermore, these maps may use different modalities, such as Synthetic Aperture Radar (SAR), which differ significantly in appearance from RGB photos. Existing retrieval-based methods are often too slow for real-time deployment or fail to handle these massive resolution discrepancies.
Methodology: Unsupervised Cross-Scale Learning
The proposed framework employs a dual-encoder architecture trained through unsupervised contrastive learning to align visual representations across scales.
Dual-Encoder Architecture
- Map Encoder ($f$): A pixels-to-pixels Fully Convolutional Network (FCN-ResNet50) that processes a low-resolution map patch into a "belief map." Each pixel in this map represents a probability distribution over learned terrain categories.
- Image Encoder ($g$): A ResNet18-based encoder that compresses high-resolution camera observations into a 1D feature vector aligned with the map's terrain dimensions.
Contrastive Learning and Similarity
The system is trained by maximizing the similarity between a camera observation and its corresponding "anchor" pixel in the map, while minimizing similarity with other images in the batch. Instead of standard cosine similarity, the authors utilize the Bhattacharyya coefficient, which is better suited for comparing the probabilistic outputs of the encoders.
Probabilistic Inference
The output is modeled as a Dirichlet distribution. This probabilistic formulation allows the localization data to be integrated directly into standard robotics frameworks, such as Particle Filters or Factor Graphs, to provide continuous path correction.
Experimental Evaluation
The framework was evaluated on two distinct "cross-scale" datasets:
- Scott Reef Dataset (Underwater): Matches high-resolution seafloor imagery to an 8x lower-resolution RGB reconstruction map.
- Kempten Dataset (Aerial): Matches RGB satellite observations to Sentinel-1 SAR maps with a massive 70x scale difference.
Key Results and Insights
- Scale and Modality Invariance: The model successfully learned to map blurry SAR pixels and detailed RGB photos to a shared representation space. This allows the system to recognize that a low-resolution radar signature and a high-resolution photo represent the same terrain type (e.g., forest or water).
- Real-Time Efficiency: By utilizing a pixels-to-pixels map encoder, the system avoids slow sliding-window searches. On a Jetson Nano, the map encoding process took approximately 54 seconds, compared to over 1,000 seconds for competing architectures like CVM-Net.
- Emergent Semantics: Despite being unsupervised, the learned representations naturally cluster into recognizable terrain categories, effectively performing unsupervised semantic segmentation within the belief maps.
- Drift Correction: Integration with a particle filter reduced localization errors by 15% to 60%, demonstrating the system's ability to effectively mitigate dead-reckoning drift.
Conclusion and Future Work
This research demonstrates that unsupervised contrastive learning can bridge the gap between high-resolution robot observations and low-resolution geospatial maps. While the current system is efficient, future work will address bandwidth constraints for underwater communication by exploring map compression techniques. Additionally, researchers aim to improve robustness in dynamic environments where terrain features may change over time.
Summary updated