Semi-Autonomous Control of Unmanned Ground Vehicles

Teleoperating a ground robot is very difficult. Humans leverage substantial proprioceptive cues to estimate safe travel speeds and compensate for motion jitter in their visual cortex. To address this problem, MIC proposes develop a semi-autonomy system able to lead or follow the operator. The system will support long-distance scene understanding use a single color camera, beyond the range of traditional stereo, to provide highly capable operations in complex terrain. Relying on monocular vision has the advantage of being a passive sensor, which is better suited for tactical missions than LADAR. We propose three complementary research thrusts that begin to interpret the scene. First, categorizing terrain traversability based on appearance and motion allows extrapolating nearby, highly accurate estimates into the far scene. Second, classifying objects based on their shape and texture allows estimating their size based on heuristics in a database, and thus estimating their distance. Finally, scene affordances with respect to navigation will be leveraged by learning mappings of visual shape categories to navigation strategies. The information from these three perceptual processes will be aggregated into a symbolic map, and used by a planner to create routes through the environment.