Evaluation of Confidence-Aware Prediction of Uneven Terrain Risk for Self-Supervised Learning by a Mobile Robot

For autonomous navigation of mobile robot in uneven outdoor terrain environments, recognition of terrain traversability is still a challenging problem. One promising approach to this problem is self-supervised learning, which relies only on the robot sensor information both for exteroceptive sensing as input and proprioceptive sensing as output. In order to reduce the cost to run over various terrains by the robot, efficient sampling of the training dataset for the terrain traversability prediction is of great importance. Toward this end, this paper presents an evaluation of confidence-aware prediction framework of terrain traversing risk. A CNN model with Monte Carlo dropout was implemented for the predictor. In the experiment, it was confirmed that the traversing cost can be predicted well for various uneven natural terrains. The confidence-aware approach was evaluated in the experiment and some future directions to utilize the confidence were discussed.