A Model Attention and Selection Framework for Estimation of Many Variables, with Applications to Estimating Object States in Large Spatial Environments

Robots performing service tasks such as cooking and cleaning in human-centric environments require knowledge of certain environmental states in order to complete tasks successfully. For example, storage locations of specific ingredients and utensils are needed for cooking; dirtiness of particular regions of space may be required for efficient cleaning. Typically these task-critical states cannot be directly observed, and must be estimated by using (noisy) perception and prior domain knowledge. Bayesian filtering solves such estimation problems for a wide variety of state characteristics: given a particular set of variables (uncertain states) to be estimated, Bayesian filtering techniques most likely already exist in that particular regime. While much effort has gone into developing various estimators, less attention has been placed on why the particular estimation problem arises. In this work, I argue that state estimation should no longer be treated as a black box. Estimating large sets of variables is computationally costly; just because a technique exists to estimate the values of certain variables does not justify its application. For robots whose ultimate mission is to complete tasks, only variables that are relevant to successful completion should be estimated. Returning to cooking and cleaning, while cooking, a robot should not prioritize estimating cleanliness of its surroundings. Similarly, while cleaning a specific room, not only should a robot not be concerned with estimating variables used in the cooking task, it should not even estimate cleanliness of other rooms. Of course, the selection of relevant variables is not so clear-cut in practice. Lack of cleanliness in the kitchen environment may lead to food contamination during cooking. Yet, as argued earlier, we want to avoid estimating all uncertain variables at once. Instead, I propose to initially only track a minimal set of directly-relevant variables, and gradually increase the sophistication of models on-demand, in a local fashion. This estimator refinement process is triggered by violations in expectations of task success. With respect to state estimation, if observed empirical quantities differ significantly from the current probabilistic model, then this indicates the model must be improved. In the remainder, I demonstrate this through a proof-of-concept case study.