Silvia Izquierdo Badiola, Guillem Alenyà, Carlos Rizzo
One of the main challenges for successful human-robot
collaborative applications lies in adapting the plan to the
human agent's changing state and preferences. A promising
solution is to bridge the gap between agent modelling and
AI task planning, which can be done by integrating the
agent state as action costs in the task planning domain.
This allows for the plan to be adapted to different
partners, by influencing the action allocation. The
difficulty then lies in setting appropriate action costs.
This paper presents a novel framework to learn a set of
planning action costs considering the preferred actions for
an agent based on their state. An evolutionary optimisation
algorithm is used for this purpose, and an action outcome
simulator is developed to act as the black-box function,
based on both an agent model and an action type model. This
addresses the challenge of collecting data in HRC
real-world scenarios, accelerating the learning for
posterior fine-tuning in real applications. The coherence
of the models and the simulator is proven through a
conducted survey, and the learning algorithm is shown to
learn appropriate action costs, producing plans that
satisfy both the agents' preferences and the prioritised
plan requisites. The resulting system is a generic
learning framework integrating components that can be
easily extended to a wide range of applications, models and planning formalisms.
Paper: Open pdf file.
The following video describes the important concepts of the paper, as presented at RO-MAN 2023.
The survey designed to evaluate the validity of the action outcome simulator can be found in this link: Try the survey!.