Forecasting off-road trafficability from terrain appearance

When people drive off-road, we look at the upcoming terrain and make a variety of judgments-whether to avoid or attempt to cross a patch of terrain, whether to slow down or speed up, etc. We consider a variety of factors including perceived slope, obstacles, resistance, traction, sinkage, roughness, and the limitations of our perception. We judge many of the handling factors based our recollections of driving on other terrain with similar visual appearance. Perception and terrain understanding algorithms with similar capabilities are needed for unmanned ground vehicle (UGV) autonomous mobility. The objective of this research is to begin to develop methods that can be used by a UGV to learn to associate terrain appearance with handling on the terrain. We demonstrate methods to identify models of speed and acceleration as function of throttle command, and power consumption as a function of speed and acceleration using data collected by on-board sensors as the UGV executes test maneuvers. We demonstrate methods to characterize terrain type from visual appearance. We investigate the hypothesis that terrain with different handling characteristics can be discriminated based on visual appearance characterization.