Sensing Anomalies as Potential Hazards: Datasets and Benchmarks

Many emerging applications involve a robot operating autonomously in an unknown environment; the environment may include hazards, i.e., locations that might disrupt the robot operation, possibly causing it to crash, get stuck, and more generally to fail its mission. Robots are usually capable to perceive hazards that are expected during system development and therefore can be explicitly accounted for when designing the perception subsystem. For example, ground robots can typically perceive and avoid obstacles or uneven ground. In this paper, we study how to provide robots with a different capability: detecting unexpected hazards, potentially very rare, that were not explicitly considered during system design. Because we don't have any model of how these hazards appear, we consider anything that is novel or unusual as a potential hazard to be avoided.