A Unified Approach to Sensor-Based Terrain Characterization and UGV Mobility Prediction

Unmanned ground vehicles (UGVs) will play an important role in the nation’s next-generation ground forces. One key limitation on autonomous terrain analysis arises from current approaches to sensor data processing. A significant amount of research in the robotics community has been devoted to interpreting data from common UGV sensors. Most current approaches to non-geometric terrain property analysis rely on classification: that is, analyzing remotely observable “features” (such as color, texture, and/or elevation) to assign a terrain region as a member of a pre-specified semantic class, such as “rock,” “loam,” “asphalt,” etc. The purpose of this STTR research program is to develop a method for terrain characterization for UGVs operating in outdoor, unstructured environments. The methods, which showed substantial promise in Phase 1, generalize locally-sensed physical terrain features to remotely-sensed data to infer properties about UGV mobility through its surroundings. In the current concept, “local” (i.e. proprioceptive) sensors measure signals related to physical UGV-terrain interaction, including wheel torque, sinkage, angular velocity, and/or vibration signatures. Local sensor feedback is analyzed to identify terrain patches that possess unique mobility characteristics, and visual features associated with these terrain patches are correlated with remote data, thereby creating a “mobility map” of the surrounding environment. Contrary to classical terrain classification methods, this map will not delineate semantically-labeled terrain boundaries (i.e. “rock,” “loam,” “asphalt,” etc.), but rather would delineate mobility-labeled terrain boundaries (i.e. “easily traversable terrain,” “poorly traversable terrain,” etc.). The continuing proposed work would be performed as a collaboration between researchers at the Massachusetts Institute of Technology’s Robotics Mobility Group, and Quantum Signal, LLC (QS).