Reactive Combination of Belief Over Time Using Direct Perception

One issue for autonomous mobile robots operating in unknown, or partially known, domains is how to handle uncertainty in their sensor observations over time. Methods such as probablistic belief networks and survivor functions are generally unsatisfactory because they require explicit models of the robot's interactions with its environment, including possible contravening events. This information is difficult to obtain, and is philosophically incompatible with reactive behaviors. This paper presents an approach which eliminates the need for explicit models and reasoning; instead, it relies solely on directly perceivable attributes of the robot, object, and environment. The attributes qualitatively rate whether the robot's current observations are from an inherently more informed state than previous readings (e.g., from a better viewpoint). Observations from more informed states have different rates for the accrual and attrition of belief than those taken from less informed states. This paper describes the implementation, focusing on how the information state is computed using fuzzy logic, and how the state dynamically adapts a variation of Dempster's rule to generate the total belief. Data from a mobile robot tracking an unknown object demonstrates that the reactive computation of belief over time performs well for six canonical accrual and attrition cases.