Optimizing Sequences of Probabilistic Manipulation Skills Learned from Demonstration

While manipulation skills such as picking, inserting and placing were hard coded in classical setups, it is now widely understood that this leads to poor flexibility and that more general skill formulations are required to ensure re-usability in new scenarios. We thus adopt a skill-centric approach where each skill is learned independently under various scenarios but not attached to any specific task. Afterwards, complex manipulation tasks can be achieved by composing these skills in sequence or parallel. One essential challenge there is to optimize the parameters of each skill such that the success rate of the whole task is maximized. Common approaches require first a discretization of the state or action space to generate such parameters and second a precise simulator to evaluate the performances under different parameters. Instead, we propose to learn task-parameterized models of each skill directly from few human demonstrations. Such models allow us to infer the success rate of executing a skill within a new scenario conveniently, via computing a novel measure of execution confidence. This measure encapsulates both the robot state and the workspace configuration. Furthermore, we introduce task-parameterized transition skills that change the object poses of interest via translation and rotation. We show that such skills can be extremely useful for changing skill parameters and thus potentially improving the success rate of a given task. The proposed scheme optimizes skill parameters in the continuous domain without the need for simulators. We demonstrate the proposed approach on a 7 DoF robot arm solving various manipulation tasks.