Publikation
Development of Language and Task Models for Skill-based Selection and Operation of Robots using Speech Dialogue
Julian Wolter
Mastersthesis, Universität des Saarlandes, DFKI, 7/2020.
Zusammenfassung
Industry 4.0 is happening. Robots are taking over more and more tasks, even in industries
in which they have not yet been present. Products are becoming increasingly individual
and are no longer mass products off the shelf, leading to a more customized production
process. To produce high-quality products efficiently, human-machine collaboration
is often essential, in particular the transfer of control from the human to the machine.
Therefore, there is a continually growing need for intuitive and flexible control of robots
without extensive training of the worker.
In my thesis, I develop an intuitive system for robot control with the help of natural
language understanding (NLU) tools. For this purpose, I first abstract and summarize
the many different abilities of various robots by creating a skill model. In addition, a task
model defines which tasks can be solved and how. There are often multiple different
ways to approach a task, and the most suitable alternative has to be chosen depending
on the capabilities of the available robots. The task model can later be mapped to real
robots using my skill model. Besides, objects in the environment are modelled and given
properties to allow for interaction with them. Once the tasks, robots and objects are
defined, the system creates a training corpus to train the NLU system and make it ready
for use. An appropriate dialogue model now makes it possible to trigger a task intuitively
by voice. To put it all together, I show in my work how to assign the generated task to the
available robots with the use of the skill model. Finally, my system sends out high-level
robot commands to fulfil the request of the user.
For evaluation purposes, I modelled the MRK4.0 laboratory in my work and showed
some possible applications. The model allows representing all scenarios successfully, and
the system distributed the tasks sensibly. I evaluated the intuitiveness of the language
component with an exploratory user study. For this purpose, I created a simulated
environment in which participants in the study should solve some predefined jobs.
There was no training in advance. Most participants could solve all tasks after a small
explorative phase in the beginning. In both cases, the system fully met expectations.