Abstract
In off-road applications, where mobile robots operate on rough environments, the physical properties of the terrain play a key role on their performance. An extreme example is posed by planetary rover missions to Mars, for which communication constraints and the inability of vision-based approaches to detect non-geometric hazards, e.g. sand traps hidden below thin duricrusts, can lead to permanent immobilisation, as experienced by NASA's Spirit rover. To prevent such events, this paper proposes a method to classify dry granular soils according to their physical properties by using an on-board sensor system for on-line analysis of sinkage, slippage and vibrations of the hybrid wheel-legs mounted on a highly mobile robot. As reflected by the experimental results, obtained using a single wheel-leg test bed, the novel approach produces an efficient and robust differentiation of soils with dissimilar physical properties. This output can enable autonomous avoidance of non-geometric hazards without endangering the mobility of the mission. Different classifier algorithms are trained, validated and compared in terms of classification accuracy and computational efficiency, revealing the advantages and disadvantages of each approach