The odd one out task: Toward an intelligence test for robots

Abstract

Detecting the outlier in a set of objects is a fundamental task used in a wide variety of intelligence tests. This paper proposes a theoretical model that allows a robot to interactively estimate the pairwise similarity between everyday objects and use this knowledge to solve the odd one out task. That is, given a set of objects, the robot's task is to select the one object that does not belong in the group. In our experiments, the robot interacted with fifty different household objects (by applying five different exploratory behaviors on them) and perceived auditory and proprioceptive sensory feedback. Pairwise object similarity was estimated for different behavior and modality contexts. In a series of subsequent tests, three objects from a given category (e.g., three cups or three pop cans) along with one object from outside that category were selected and the robot's internal models were queried to pick the object that does not belong in the group. The object similarity relations learned by the robot were used to pick the most dissimilar object, with success rates varying from 45% to 100%, depending on the category. The results show that the learned similarity measures were sufficient to capture some of the common properties of human-defined object categories, such as cups, bottles, and pop cans.