For robots to build trustable interactions with users two aspects will be crucial during the next decade. First, the ability to produce explainable decisions combining reasons from all the levels of the robotic architecture from low to high level; and second, to be able to effectively communicate such decisions and re-plan according to new user inputs in real-time along with the execution.
COHERENT will develop a novel framework to coordinate explanations originated at the different robotic levels and to be able to deliver these explanations during the execution of the task. To have effective interactions, an interface of communication with the user will be developed to both explain and receive inputs in the form of user preferences, requirements or suggestions to execute a task, at different levels of human expertise. Validation will entail a new benchmark to assess acceptance and effectiveness of explanations based on experiments with subjects.
We will demonstrate our framework for hierarchical explanation components through a manipulation task of assisting a human to fold clothes. Cloth manipulation requires considering bi-manual manipulations, environmental constraints, and perception of textiles for its state estimation. We will build on previous results on cloth manipulation to develop explainable machine learning techniques from the perception, learned movements, task planning and interaction layers, based on a generic state-and-transitions representation that is shared across the layers. The COHERENT framework will be integrated into the standard planning system ROSplan.
The interface for HRI will build on the idea of a hierarchy of layers from robotics to build different technical content explanations depending on user expertise. We will define an interface to interchange information with the user in both directions.
Robotics is one of the applications of AI systems that will interact in close relationship with humans, and that will require the highest social acceptance to be able to work in collaboration both in industrial and domestic environments. This proposal takes the challenge of designing an explainable decision-making system tailored to robotic applications in assistive tasks.
To do that, machine learning techniques applied for manipulation tasks will be expanded to build explainable systems, and the explainability at the different robotic layers will enable a systematic method to define different levels of technical knowledge depending on stakeholders’ previous knowledge.
Finally, one of the objectives of the COHERENT project is to establish measures for the effectiveness of explanations through user studies and define assistive tasks to test our explainable methods creating clear and reproducible protocols.
Start date: (36 months)
Funding support: 562 736 €