We define the problem of visuomotor coordination and identify bottleneck problems in the implementation of general purpose vision and action systems. We conjecture that machine learning methods provide a general purpose mechanism for combining specific visual and action modules in a task-independent way. We also maintain that successful learning systems reflect realities of the environment, exploit context information, and identify limitations in perceptual algorithms which cannot be captured by the designer. We then propose a multi-step find-and-fetch mobile robot search and retrieval task. This task illustrates where current learning approaches provide solutions and where future research opportunities exist.