A Bayesian Approach to Terrain Map Inference based on Vibration Features

Abstract

In this paper, we represent a terrain inference method based on vibration features. Autonomous navigation in unstructured environments is a challenging problem. Especially, the detailed interpretation of terrain in unstructured environments is necessary to set an efficient navigation trajectory. As the vibration features are obtained from interactions between the robot and terrain, terrain inference based on vibration can be conducted. To perform the terrain inference for robot path and unobserved field simultaneously, we use a Bayesian random field for structured prediction method. The robot path and the unobserved field are represented by the Conditional Random Field (CRF), and based on the terrain information observed on the robot path, the terrain of the region that the robot does not approach is estimated together. The proposed algorithm is tested with a 4WD mobile robot and real-terrain testbed.