EVC-Mamba corrects IMU drift without GPS using AI velocity sensor

Accurate vehicle localization in tunnels, garages, or other GNSS-denied environments remains a critical challenge for autonomous driving. Traditional inertial navigation systems (INS) accumulate unbounded drift without periodic GPS corrections, forcing reliance on costly external sensors like wheel encoders or Doppler radar. Now, a team from TU Munich has developed EVC-Mamba (Evidential Velocity Correction using Mamba), a deep learning architecture that transforms existing onboard sensor data—such as wheel speeds and steering angles—into a virtual velocity measurement, eliminating the need for any extra hardware.

The system leverages Mamba, a recent selective state space model architecture, to capture complex temporal dynamics of vehicle motion more efficiently than traditional RNNs or Transformers. Additionally, evidential deep learning with a Normal-Inverse-Gamma distribution provides principled uncertainty estimates for each velocity prediction. These uncertainty-aware estimates are fed as corrective measurements into an Error-State Extended Kalman Filter to minimize position drift. On real-world vehicle data across various GNSS outage durations, EVC-Mamba maintained localization accuracy within 10% of a dedicated external velocity sensor. The entire pipeline runs at 40 Hz on embedded edge hardware, making it practical for real-time deployment.