Solving Motion Tasks with Challenging Dynamics by Combining Kinodynamic Motion Planning and Iterative Learning Control

This work considers the problem of robots with challenging dynamics having to solve motion tasks that consist in transitioning from an initial state to a goal state in an environment that is obstructed by obstacles. We propose a novel combination of methods from motion planning and iterative learning control to solve these motion tasks. The proposed method only requires an approximate, linear model of the nonlinear, possibly underactuated robot dynamics. The proposed method employs the approximate, linear model in a kinodynamic rapidly exploring random tree to plan a state trajectory that solves the motion task. Based on the distance to the obstacles, the most relevant samples of the planned trajectory are selected as reference points. Lastly, point-to-point iterative learning control is employed to learn a feedforward input trajectory that leads to the state trajectory precisely tracking the reference points despite the robot's nonlinear real-world dynamics. The proposed method is validated in real-world experiments on a two-wheeled inverted pendulum robot that has to solve a motion task that requires the robot to perform an agile motion to dive beneath an obstacle.