VIO-GO

Sakhrieh 2025 · Paper

One-line summary — VIO-GO wraps an event-based VIO pipeline (Ultimate SLAM) in a gradient-based parameter optimizer: Batch Gradient Descent with momentum automatically tunes front-end and back-end parameters per scenario, cutting mean position error by 60% versus fixed-parameter operation in HDR and dynamic scenes.

Problem

Event-based VIO systems that represent events as motion-compensated edge images (Ultimate SLAM and its lineage) are notoriously sensitive to configuration: the number of events per frame, noise-rate thresholds, feature-detector settings, and back-end keyframe counts all depend on scene texture, event rate, and sensor unit. Per-sequence manual tuning produces the published headline numbers but is impractical in the field — exactly the Industry 4.0 warehouse/logistics settings with low light and fast motion where event cameras are most attractive. VIO-GO targets this quieter failure mode: not a new estimator, but automated tuning of the one we have.

Method & architecture

Results

Why it matters for SLAM

Most event-VIO papers introduce new estimators; VIO-GO instead addresses the quieter reason such systems underperform in practice — poorly matched parameters for the scenario at hand — and shows that automated tuning can recover (and in HDR scenes exceed) hand-tuned performance. For a practitioner deploying event-based VIO, it is a useful reference for which knobs dominate performance (event window size first). It sits at the applied end of the event-SLAM literature, complementing the architectural advances of ESVIO and DEVO.