Autonomously improving binocular depth estimation

We investigate how an autonomous humanoid robot with an initially inaccurate binocular vision system can learn to correct inconsistencies in its understanding of distance using information and resources that might be available to a human infant. We defined a consistent depth estimator as a Euclidean distance metric where the unit of measurement is determined autonomously. We found that an error signal that exploits actions that maintain invariant distance is a powerful tool for correcting inconsistency. Our results show that a heuristic search algorithm, run incrementally as new data become available, can efficiently (i.e. with few samples) correct inconsistencies and improve depth estimates. Keywords—Binocular Vision, Depth Estimation, Active Learning, Autonomous Robot