Terrain Sensing for Mobile Robots

For mobile robots in rough terrain, the ability to safely traverse terrain is highly dependent on mechanical properties of that terrain. For example, a robot may be able to climb up a rocky slope with ease, but slide down a sandy slope the same grade. With mobile robots being employed for planetary exploration and UGVs being developed for missions on Earth, the ability to predict these mechanical terrain properties from a distance is becoming increasingly important.

This project focuses on:

(*) Classifying natural terrain based on visual features, such as color, visual texture, and range data,
(*) Learning visual classification on-line, so that a robot can improve its terrain recognition based on its experiences,
(*) Autonomously identifying mechanically-distinct terrain classes to eliminate the need for human supervision in establishing the list of terrain classes, and
(*) Estimating the mechanical terrain properties associated with each of the terrain classes.

The goal of this work is to be able to set a robot down in a previously unexplored environment, and after driving around for a short period of time have it be able to look out in the distance and predict mechanical properties of the terrain it sees.

Experiments for this project have been performed using a four-wheeled mobile robot in natural outdoor terrain.