Learning risk-mediated traversability maps in unstructured terrains navigation through robot-oriented models

In robotic navigation, safety and efficiency play an important role and must be evaluated together. This paper proposes a simple and efficient method to derive the traversability maps of unstructured environments and the optimal path to be followed to reach target locations based on the specific characteristics of different types of robots available. The optimal solution minimises the path length while maintaining the risk associated with that path below the maximum acceptable upper bound. The ability of each robot to traverse terrains with specific characteristics is formalised and modelled using simple and efficient neural networks and trained in a dynamic simulation environment. The proposed shallow network topology achieves results, in terms of accuracy, that are comparable with other standard classifiers and more complex deep networks. Applying this procedure to different robotic structures, the best system within the team (wheeled, legged, and hybrid) can be selected to accomplish a specific assigned task. The proposed strategy, together with the obtained simulation results is presented, carefully analysed, and then compared using real-life simulated scenarios.