Estimating terrain parameters for a rigid wheeled rover using neural networks

• Terrain parameter estimation using neural networks is proposed.
• Preliminary results from micro-rover instrumented mobility system tests.
• On-line terrain parameter estimation to avoid immobilization.

This paper presents a method for extracting data on regolith online with a planetary exploration micro-rover. The method uses a trained neural network to map engineering data from an instrumented chassis to estimates of regolith parameters. The target application for this method is a low-cost micro-rover scout on Mars that will autonomously traverse the surface and detect changes in the regolith cohesion and shearing resistance without the need for dedicated visual sinkage estimation on each wheel. This method has been applied to Kapvik, a low-cost 30 kg micro-rover analogue designed and built for the Canadian Space Agency. Data was collected using a motor controller interface designed for Kapvik using off-the-shelf components. The neural network was trained from parameters derived by classical terramechanics theory using Matlab’s Neural Network Toolbox. The results demonstrate a proof of concept that neural networks can estimate the terrain parameters which may have applications for automated online traction control.