Teaching Robots Behaviors Using Spoken Language in Rich and Open Scenarios

Overview & Synopsis

Social robots like Pepper are already found "in the wild". Their behaviors must be adapted for each use case by experts. Enabling the general public to teach new behaviors to robots may lead to better adaptation at lesser cost. In this thesis, we study a cognitive system and a set of robotic behaviors allowing home users of Pepper robots to teach new behaviors as a composition of existing behaviors, using solely the spoken language. Homes are open worlds and are unpredictable. In open scenarios, a home social robot should learn about its environment. The purpose of such a robot is not restricted to learning new behaviors or about the environment: it should provide entertainment or utility, and therefore support rich scenarios. We demonstrate the teaching of behaviors in these unique conditions: the teaching is achieved by the spoken language on Pepper robots deployed in homes, with no extra device and using its standard system, in a rich and open scenario. Using automatic speech transcription and natural language processing, our system recognizes unpredicted teachings of new behaviors, and a explicit requests to perform them. The new behaviors may invoke existing behaviors parametrized with objects learned in other contexts, and may be defined as parametric. Through experiments of growing complexity, we show conflicts between behaviors in rich scenarios, and propose a solution based on symbolic task planning and priorization rules to resolve them. The results rely on qualitative and quantitative analysis and highlight the limitations of our solution, but also the new applications it enables.