Trajectory optimization and reinforcement learning are combined for versatile and robust perceptive legged locomotion.