Efficient Imitation Learning with Local Trajectory Optimization