FusionLoc: Camera-2D LiDAR Fusion Using Multi-Head Self-Attention for End-to-End Serving Robot Relocalization
It addresses a practical issue in operating serving robots, where pose estimation failures require manual intervention, by improving end-to-end relocalization, though it appears incremental as it builds on existing sensor fusion approaches.
The paper tackles the problem of serving robot relocalization by proposing FusionLoc, a deep neural network that fuses camera and 2D LiDAR data using multi-head self-attention to directly predict robot pose from onboard sensors, achieving better performance than previous single-sensor or simple fusion methods.
As technology advances in autonomous mobile robots, mobile service robots have been actively used more and more for various purposes. Especially, serving robots have been not surprising products anymore since the COVID-19 pandemic. One of the practical problems in operating a serving robot is that it often fails to estimate its pose on a map that it moves around. Whenever the failure happens, servers should bring the serving robot to its initial location and reboot it manually. In this paper, we focus on end-to-end relocalization of serving robots to address the problem. It is to predict robot pose directly from only the onboard sensor data using neural networks. In particular, we propose a deep neural network architecture for the relocalization based on camera-2D LiDAR sensor fusion. We call the proposed method FusionLoc. In the proposed method, the multi-head self-attention complements different types of information captured by the two sensors to regress the robot pose. Our experiments on a dataset collected by a commercial serving robot demonstrate that FusionLoc can provide better performances than previous end-to-end relocalization methods taking only a single image or a 2D LiDAR point cloud as well as a straightforward fusion method concatenating their features.