Intelligence test scores have long been shown to correlate with a wide variety of other abilities. The goal of this thesis is to enable a robot to solve some of the common tasks from intelligence tests with the intent of improving its performance on other real-world tasks. In other words, the goal of this thesis is to make robots more intelligent. We used an upper-torso humanoid robot to solve three common perceptual reasoning tasks: the object pairing task, the order completion task, and the matrix completion task. Each task consisted of a set of objects arranged in a specific configuration. The robot's job was to select the correct solution from a set of candidate solutions.

To find the solution, the robot first performed a set of stereotyped exploratory behaviors on each object, while recording from its auditory, proprioceptive, and visual sensory modalities. It used this information to compute a set of similarity scores between every pair of objects. Given these similarity scores, the robot was able to deduce patterns in the arrangement of the objects, which enabled it to solve the given task. The robot repeated this process for all the tasks that we presented to it. We found that the robot was able to solve all the different types of tasks with a high degree of accuracy.

There have been previous computational solutions to tasks from intelligence tests, but no solutions thus far have used a robot. This thesis is the first work to attempt to solve tasks from intelligence tests using an embodied approach. We identified a framework for solving perceptual reasoning tasks, and we showed that it can be successfully used to solve a variety of such tasks. Due to the strong correlation between intelligence test scores and performance in real-world environments, this suggests that an embodied approach to learning can be very useful for solving a wide variety of tasks from real-world environments in addition to tasks from intelligence tests.