Concept Evaluation of Modeling Terrain Mechanics by a Neural Network

Abstract :

In the field of terrain mechanics the Bekker theory is mostly used for describing the interaction between a wheeled or tracked vehicle and the terrain. No application is known where the Bekker theory is applied to legged robots. Furthermore, the most legged robot simulators use simple coulomb friction models to simulate the interaction of the feet with the ground. Only a few papers deal with contact models of sandy terrain and legged robots. In these papers molecular dynamics approaches are described, which are far away from real-time capabilities. The contribution of this paper is an investigation of a new concept with the goal to evaluate the purpose of a neural network as model for contact dynamics. Neural networks are already successfully used for estimating the behaviour of a wheeled vehicle on sloping terrain. Two real-world experiments are performed and the results are used to learn the parameters of the neural network. The first experiment measures the acceleration and the impact depth of a cylinder, while falling into sand (or into a regolith simulating material). The second experiment measures the force needed to move a cylinder that is dipped into sand with one side and that is mounted to a linear bearing with the other ending. The experiments are done with different parameters of the cylinders (mass, size, depth, etc.). Half of both experiment results are used to learn the neural network parameters and the other results are used to evaluate the learned network. Finally, the usability of this approach and possible future research will be discussed in this paper.