To address how the highly stereotyped retinotectal pathway develops in zebrafish, we used fixed-tissue and time-lapse imaging to analyze morphology and behavior of wild-type and mutant retinal growth cones. Wild-type growth cones increase in complexity and pause at the midline. Intriguingly, they make occasional ipsilateral projections and other pathfinding errors, which are always eventually corrected. In the astray/robo2 mutant, growth cones are larger and more complex than wild-type. astray axons make midline errors not seen in wild-type, as well as errors both before and after the midline. astray errors are rarely corrected. The presumed Robo ligands Slit2 and Slit3 are expressed near the pathway in patterns consistent with their mediating pathfinding through Robo2. Thus, Robo2 does not control midline crossing of retinal axons, but rather shapes their pathway, by both preventing and correcting pathfinding errors.