Uncertainty-Aware Velocity Correction for Proprioceptive Vehicle Localization using Evidential Mamba

Technical Summary: Uncertainty-Aware Velocity Correction for Proprioceptive Vehicle Localization using Evidential Mamba

Problem Statement
Reliable vehicle localization in Global Navigation Satellite System (GNSS)-denied environments (e.g., tunnels, parking structures) remains a critical challenge. While Inertial Navigation Systems (INS) are standard, they suffer from unbounded position drift due to the accumulation of noise and bias in Inertial Measurement Unit (IMU) data without external corrections. Existing solutions often rely on dedicated infrastructure, expensive external sensors (e.g., Correvit), or complex multi-sensor fusion, which introduce deployment barriers. Furthermore, while data-driven methods have improved IMU-based dead reckoning, their effectiveness is often limited to short outage durations, and many lack principled uncertainty quantification, which is essential for robust state estimation.

Methodology
The paper proposes Evidential Velocity Correction using Mamba (EVC-Mamba), a learning-based architecture that transforms onboard vehicle sensor data into a "virtual velocity sensor" to correct IMU drift without additional hardware. The framework consists of three core components:

1. Mamba-Based Temporal Modeling:
   Instead of using Recurrent Neural Networks (RNNs) or Transformers, the architecture employs a Mamba-based Selective State Space Model (SSM). This model captures the temporal dynamics of vehicle motion from onboard sensor sequences (steering wheel angle, wheel speeds, lateral acceleration, brake pressure, etc.). The Mamba architecture offers linear-time inference complexity, making it suitable for real-time deployment, while its input-dependent parameters allow it to adapt to varying driving conditions.

2. Evidential Deep Learning for Uncertainty Quantification:
   To address the inherent uncertainty in estimating velocity from kinematic sensors, the model utilizes evidential deep learning. Rather than predicting a single point estimate, the network outputs the parameters of a Normal-Inverse-Gamma (NIG) distribution. This allows the model to jointly predict the mean velocity ($\tilde{v}_t$) and its associated uncertainty ($\delta\tilde{v}$), separating aleatoric (data) and epistemic (model) uncertainty. This is achieved in a single forward pass, avoiding the computational cost of Monte Carlo Dropout or ensemble methods.

3. ES-EKF Integration:
   The predicted velocity and its uncertainty are fused into an Error-State Extended Kalman Filter (ES-EKF) as a virtual measurement. The filter uses the uncertainty estimates to weight the correction signal, effectively constraining the growth of position and velocity errors during GNSS outages. The system operates on a state vector comprising position, velocity, orientation, and sensor biases.

Key Contributions

• GNSS-Independent Velocity Correction: A learning-based virtual velocity sensor that corrects IMU drift using only standard onboard sensors (steering angle, wheel speeds), eliminating the need for external hardware or infrastructure.
• Uncertainty-Aware Estimation: The integration of a Mamba-SSM with an evidential framework to provide principled, single-pass uncertainty quantification alongside velocity estimates, enabling robust correction signals for state estimation.
• Real-Time Edge Deployment: The architecture achieves real-time performance (40 Hz) on edge hardware (NVIDIA Orin) with low computational cost (22.8 MFLOPs), outperforming Transformer-based baselines in efficiency while maintaining superior accuracy.
• ES-EKF Integration: A seamless integration of the learned virtual measurements into a standard Error-State EKF for inertial navigation.

Results
Evaluated on the ReV-StED dataset (real-world vehicle data with ground truth from RTK/INS), the system demonstrates:

• Velocity Estimation Accuracy: EVC-Mamba outperforms both kinematic baselines and Transformer-based architectures, achieving a lateral velocity RMSE of 0.011 m/s (compared to 0.017 m/s for Transformers).
• Uncertainty Calibration: The predicted uncertainties are well-calibrated, closely tracking the empirical coverage of the true errors.
• Localization Performance: During GNSS outages, EVC-Mamba significantly reduces position drift compared to uncorrected IMU dead reckoning.
  • In a 10-minute outage, uncorrected IMU drift reached 142 meters, whereas EVC-Mamba reduced this to 5.9 meters.
  • The performance is within 10% of a dedicated external hardware sensor (Correvit), which serves as the practical upper bound for this task.
• Efficiency: The model operates at 40 Hz with 20–24 ms latency, making it roughly three times more computationally efficient than Transformer baselines.

Significance and Claims
The authors claim that EVC-Mamba provides a cost-efficient, hardware-free alternative for reliable proprioceptive localization during prolonged GNSS outages. By leveraging existing onboard sensors and advanced sequence modeling, the system achieves localization accuracy comparable to expensive external sensors without the associated deployment costs. The work highlights the potential of combining Selective State Space Models with evidential learning to create robust, uncertainty-aware perception modules for autonomous vehicles, specifically addressing the gap between short-term data-driven corrections and long-term reliable navigation.