Improving Policy Optimization with Generalist-Specialist Learning