Robust Traversability Prediction Using Multiple Costs for Quadruped Robot in Random Terrains

The quadruped robot has to assess the feasibility of upcoming terrains before making contact to safely traverse various environments. This assessment is called traversability in the literature on quadruped robots. Trasversability has recently posed challenges due to a high-dimensional system that leads to long computational times. Furthermore, exteroceptive observations often suffer from noise that potentially causes misinterpretations of terrains and results in an inaccurate assessment. This paper proposes a robust traversability predictor to tackle these issues by utilizing a Convolutional Neural Network (CNN) encoder, CNN decoder, and Multi-layer Perceptron (MLP) to predict multiple costs associated with traversability. The integration of the CNN encoder and decoder helps mitigate the effect of noise in exteroceptive observations, while the MLP network serves as a predictor for multiple costs. The proposed method utilizes the information collected from a physics simulator to avoid hand-crafted multiple-cost labeling. It can predict a comprehensive set of costs that overcomes the limitations of relying on a single cost metric. It also achieves faster computational time by utilizing neural networks, in contrast to the model-based approach in the literature. The robustness of the proposed method is validated by comparing it to a baseline noise-free prediction model and an existing method in the literature. The results indicate that the proposed method exhibits the lowest prediction errors. Therefore, despite the noise in exteroceptive observations, the proposed multiple cost-based traversability predictor has better accuracy and robustness than the baseline and existing methods.