Generation of Optimal Walking-Like Motions Using Dynamic Models with Switches, Switch Costs, and State Jumps.