Can robots learn from humans interactively and cooperatively? In the following video, roboticist Katharina Muelling generously imparts her ping pong expertise to a Barrett WAM arm robot equipped with her team's mixture of movement primitives (MoMP) algorithm, then plays against the robot.
(Robot learns ping pong from roboticist video)
In this paper, we presented a new framework that enables a robot to learn basic cooperative table tennis. To achieve this goal, we created an initial movement library from kinesthetic teach-in and imitation learning. The movements stored in the movement library can be selected and generalized using a mixture of movement primitives algorithm. As a result, we obtain a task policy that is composed of several movement primitives weighted by their ability to generate successful movements in the given task context. These weights are computed by a gating network and can be updated autonomously.
The setup was evaluated successfully in a simulated and real table tennis environment. We showed in our experiments that both (i) selecting movement primitives from a library based on the current task context instead of using only a single demonstration and (ii) the adaptation of the selection process during a table tennis game improved the performance of the table tennis player. As a result, the system was able to perform a game of table tennis against a human opponent.
(Learning to Select and Generalize Striking Movements in Robot Table Tennis)