In our framework, we train a NeRF model to accurately represent the environment in which the UAV is flying and use the discrepancy between the UAV-captured images and NeRF-rendered scenes to compute visual corrections. These corrections are then integrated with the GNSS corrections learned by the GNN and used to predict the UAV’s absolute position. The proposed approach is validated on UAV real-world urban environment datasets with emulated GNSS pseudorange measurements. The results demonstrate that our framework improves UAV localization accuracy compared to traditional GNSS and camera-based positioning methods under different measurement noise conditions. 

Adyasha Mohanty and Grace Gao, Fusing GNN Enhanced GNSS with Neural Radiance Fields for UAV Navigation, ION International Technical Meeting (ION ITM 2024), Long Beach, CA, Jan 2024.  [paper] [slides] [video] 

Utilizing drift dynamics, we create an a priori relative position estimate, adaptively adjusting the integer search space based on noise. Our algorithm outperforms conventional methods like LAMBDA in a simulated tandem drifting scenario, achieving cm-level accuracy despite noisy GPS measurements. 

Adyasha Mohanty, Remy Zawislak, Sriramya Bhamidipati, and Grace Gao, Precise Relative Positioning for Tandem Drifting Cars, Proceedings of the Institute of Navigation GNSS+ conference (ION GNSS+ 2021), St. Louis, MO, Sep 2021. [paper] [slides] [video] 

UWBs  provide noisy inter-ranging measurements at a high sampling frequency. With these range measurements, we form an a priori estimate of the relative position that adaptively constrains the position search space, and subsequently reduces the induced integer search space. Adaptively constraining the integer search space improves integer fixing rate, enabling precise relative positioning. We demonstrate our algorithm’s performance on a real-world dataset collected with two cars in Stanford University. Our algorithm achieves improved accuracy over baseline methods, even in challenging urban settings.

Adyasha Mohanty, Asta Wu, Sriramya Bhamidipati and Grace Gao, Precise Relative Positoning via Tight Coupling of GPS Carrier Phase and Multiple UWBs, IEEE Robotics and Automation Letters.vol. 7, no. 2, pp. 5757 - 5762, Apr. 2022. DOI: 10.1109/LRA.2022.3145051 and IEEE International Conference on Robotics and Automation (ICRA) 2022  [paper] 

David Stier, Asta Wu, Adyasha Mohanty and Grace Gao, A Test Platform for UWB-Based Localization of Dynamic Multi-Agent Systems, IEEE International Conference on Robotics and Automation (ICRA) 2022.  [paper] [slides] [video] 

Reachability analysis (RA) is a technique that is commonly used to provide formal guarantees on the position under such uncertainty. In this work, we integrate a set-based RA estimation framework with nonlinear drift dynamics and double-difference GPS pseudorange measurements to provide positioning bounds for an autonomous drifting vehicle. We validate our framework on high-fidelity simulated drifting experiments with varied GPS measurement noise profiles. We show that the positioning bounds from our framework successfully bound the trajectories from a particle filter across all validation cases.

Asta Wu, Adyasha Mohanty, Anonto Zaman, and Grace Gao, Bounding GPS-Based Positioning and Navigation Uncertainty for Autonomous Drifting via Reachability, Proceedings of the Institute of Navigation GNSS+ conference (ION GNSS+ 2023), Denver, CO, Sep 2023. Best Presentation of the Session Award. [paper] [slides] [video]  

We design a particle filtering framework for joint sensor fusion and integrity monitoring of a GNSS-camera system. To account for vision faults, we derived a probability distribution over position from camera images using map-matching. We formulated a Kullback-Leibler divergence metric to assess the consistency of GNSS and camera measurements and mitigate faults during sensor fusion.

Adyasha Mohanty, Shubh Gupta and Grace Gao, A Particle Filtering Framework for Integrity Risk of GNSS-Camera Sensor Fusion, Navigation: Journal of the Institute of Navigation. Dec. 2021. doi: 10.1002/navi.455. [paper] 

We tightly couple GNSS-camera measurements within a particle filtering framework while mitigating measurement faults with an M-estimator. We also propose a data-driven method using CNNs that is fast at inference time and more robust to vision faults. We validate our framework on a real-world urban driving dataset. Our method achieves lower positioning error than baseline methods under multiple GNSS and camera measurement faults.

Adyasha Mohanty and Grace Gao, A Particle Filtering Framework for Tight GNSS-Camera Fusion using Convolutional Neural Networks, Proceedings of the Institute of Navigation GNSS+ conference (ION GNSS+ 2021), St. Louis, MO, Sep 2021. [paper][slides] [video] 

We develop a hybrid Bayesian filter that combines the tracking efficiency of a Kalman filter with the superior uncertainty modeling of a particle filter. Our filter employs Rao-Blackwellization to decouple the state into a non-linearly tracked position and linearly-tracked orientation, velocity and carrier phase integer ambiguities. This factorization allows our filter to efficiently track the state along with a multi-modal probability distribution of the position. We utilize the tracked filter probability distribution to design a metric that reliably quantifies the uncertainty in the estimated position. 

Shubh Gupta, Adyasha Mohanty and Grace Gao, Getting The Best of Particle and Kalman Filters: GNSS Sensor Fusion using Rao-Blackwellized Particle Filter, Proceedings of the Institute of Navigation GNSS+ conference (ION GNSS+ 2022), Denver, CO, Sep 2022. [paper] [slides] [video] 

Shubh Gupta, Adyasha Mohanty, and Grace Gao, Urban Localization using Robust Filtering at Multiple Linearization Points, EURASIP Journal on Advances in Signal Processing. Accepted. [paper] 

We design a hybrid framework for learning position corrections from smartphone GNSS measurements by combining a learning-based approach using a Graph Convolutional Neural Network (GCNN) with a model-based approach using a Kalman filter to predict an initial position and condition the input measurements to the GCNN. The GCNN predicts a finer position correction by applying convolution operations to the input graph, which includes satellite positions as nodes and preconditioned inputs from the Kalman filter. The GCNN can handle varying satellite visibility in urban environments and model measurements from multiple constellations and multiple signal frequencies. 

To further improve the accuracy, we have deeply coupled a Graph Neural Network (GNN) with a learnable Backpropagation Kalman Filter (BKF). This hybrid approach combines the strengths of both model-based and data-driven methods, enhancing adaptability in complex urban settings. We further augment the GNN’s measurement modeling capabilities with extended features, a novel edge creation technique, and an inductive graph learning framework. Additionally, we implement a unique backpropagation strategy that uses real-time positioning corrections to refine the performance of both the GNN and the learned Kalman filter.

Adyasha Mohanty and Grace Gao, Tightly Coupled Graph Neural Network and Kalman Filter for Improving Smartphone GNSS Positioning, Proceedings of the Institute of Navigation GNSS+ conference (ION GNSS+ 2023), Denver, CO, Sep 2023, Best Presentation of the Session Award. [paper] [slides] [video] 

Adyasha Mohanty and Grace Gao, Learning GNSS Positioning Corrections for Smartphones using Graph Convolution Neural Networks, Proceedings of the Institute of Navigation GNSS+ conference (ION GNSS+ 2022), Denver, CO, Sep 2022. [paper] [slides] [video]