New hybrid AI fusion boosts autonomous navigation accuracy 2x at 155 FPS

A new paper on arXiv from researchers Simegnew Yihunie Alaba and Yuichi Motai presents an uncertainty-aware adaptive sensor fusion framework that significantly improves pose estimation in Visual-Inertial Odometry (VIO) for autonomous navigation. The model uses a Vision Transformer (ViT) to capture temporal dependencies from IMU data and a Multiscale Convolutional Neural Network (MCNN) to learn optical flow-based motion cues from visual data. An adaptive fusion module dynamically weights these two sensor streams based on learned uncertainty estimates, making the system robust to noisy or incomplete inputs. The authors also introduce a novel uncertainty-aware loss function that explicitly incorporates prediction uncertainty during training, further improving performance in challenging conditions.

Tested on the KITTI dataset, the method outperforms existing baselines on both Absolute Trajectory Error (ATE) and Relative Pose Error (RPE) metrics, demonstrating superior localization accuracy. Notably, the entire pipeline runs at 155 frames per second on a single NVIDIA A100 GPU, making it lightweight enough for deployment on vehicles with limited computational resources. This balance of high accuracy and low latency positions the approach as a practical solution for real-world autonomous driving systems, where reliable state estimation is critical for safety in diverse environments.