https://doi.org/10.31449/inf.v45i3.3759 Informatica 45 (2021) 489–490 489
Automated Planning with Induced Qualitative Models in Dynamic Robotic
Domains
Domen Šoberl
University of Primorska, Faculty of Mathematics, Natural Sciences and Information Technologies, Koper, Slovenia
University of Ljubljana, Faculty of Computer and Information Science, Ljubljana, Slovenia
E-mail: domen.soberl@famnit.upr.si
Thesis summary
Keywords: learning qualitative models, qualitative reasoning, qualitative simulation, qualitative planning, explainable
control strategies
Received: September 22, 2021
Various methods of learning qualitative models by autonomous robots have previously been proposed as a
viable alternative to traditional numerical modeling, but an efficient way of using such models for planning
still remained an open problem. This paper summarizes a doctorat thesis, which proposes a novel domain-
independent qualitative approach to automated planning and execution of robotic plans. The proposed
method is demonstrated in five different robotic domains.
Povzetek: Razliˇ cne metode uˇ cenja kvalitativnih modelov v avtonomni robotiki so bile v preteklosti pred-
lagane kot možna alternativa tradicionalnemu numeriˇ cnemu modeliranju, toda uˇ cinkovit naˇ cin uporabe
takšnih modelov za planiranje je še vedno ostal odprt problem. Ta ˇ clanek povzema doktorska disertacijo,
ki predlaga nov, domensko neodvisen kvalitativen pristop k avtomatiziranemu planiranju in izvedbi robot-
skih planov. Disertacija predlagano metodo demonstrira v petih razliˇ cnih robotskih domenah.
1 Introduction
Cognitive robotics aims to bring robots out of their typical
industrial environments and closer to a human type of rea-
soning and cognition. Some researchers in this area have
proposed the use of qualitative modeling as a viable alter-
native to the traditional numerical modeling. Various ex-
periments showed a better training sample efficiency and
a lower susceptibility to noise when learning qualitatively.
Moreover, since qualitative models tend to be closer to hu-
man intuition, they can provide a more comprehensive ex-
planation of the learned theory.
Qualitative models may have a certain explanatory sig-
nificance to humans, but to robots, they are generally far
less useful than numerical models. To perform some task, a
robot needs to plan, and planning in continuous robotic do-
mains is typically a numerical problem. Qualitative mod-
els abstract away most if not all numerical data, so an im-
portant part of information needed by a planner is missing
from the model.
To answer the question of how qualitative models can
still assist with the robotic performance, Wiley et al. [1]
proposed to qualitatively constrain the search space of a
trial-and-error learning, thus shorten the time to find a
working solution in a numerical way. However, the ques-
tion whether qualitative models could be used to plan and
execute a robotic task without additional numerical training
still remained open.
This paper summarizes a doctoral thesis [2], which
shows that certain robotic tasks can be planned and ex-
ecuted within continuous domains using qualitative in-
formation alone. The performance can improve over
time through on-line numerical adaptation. The proposed
method was demonstrated in five different robotic domains.
2 Methodology
The thesis focuses on a specific type of qualitative mod-
els that encode information through monotonic qualita-
tive constraints (MQC), which were proposed by Benjamin
Kuipers [3]. These constraints abstract differentiable func-
tions of one or more variables to intervals of monotonically
increasing and decreasing values. A system modeled in
such a way can be simulated qualitatively using Kuipers’
program called QSIM. The thesis shows how QSIM can be
transformed into a qualitative robotic planner that can de-
vise symbolic plans based on a given qualitative model and
with some given initial and goal conditions. Such plans
determine a sequence of symbolic actions that should be
taken (e.g. accelerate, turn right, etc.), but do not deter-
mine their numerical outputs and durations.
To execute qualitative plans on a numerical system, the
thesis proposes a novel algorithm that reactively decides
the numerical outputs based on the current and past numer-
ical observations, while following a qualitative plan. Nu-
merical observations such as speeds and accelerations of
individual robot’s attributes (typically translations and ro-

490 Informatica 45 (2021) 489–490 D. Šoberl
tations of the robot or its individual parts) are used by the
algorithm to internally construct a smooth hypersurface, to
which it maps the planned qualitative states, so that the tar-
geted state is always lower on the hypersurface than the
current state. The output signals are then determined dy-
namically by following the steepest descent along the hy-
persurface. The performance of the algorithm has been
successfully evaluated for real-time execution on embed-
ded systems.
To demonstrate domain independence of the proposed
method, the implementation strictly separates the method
from the model. A new language called QDDL (Qualita-
tive Domain Description Language) is defined and used to
conduct the experiments.
3 Results
The proposed method was demonstrated with five different
robotic problems, each exposing a different aspect of qual-
itative planning and execution. Some experiments empha-
size the dynamic adaptability of qualitative models to the
given environmental parameters, while others focus more
on conceptual reasoning about the given numerical prob-
lem through qualitative planning.
Two-wheled autonomous vehicle learned a qualitative
model of differential drive by motor babbling. Using the
learned model the robot was able to pursuit a moving tar-
get and avoid collisions [4].
Pushing objects by a robot. A wheeled robot learned a
qualitative model that encoded the physics behind pushing
a rectangular box. Qualitative abstraction made the learned
model general enough to enable the robot to push boxes of
different shapes to the designated locations. [5]
Flying a quadcopter. A quadcopter autonomously
learned a qualitative model for navigating through space
and qualitatively planned and executed an escape out of an
enclosed space. [6]
Controlling the cart-pole system. The well known in-
verted pendulum problem was implemented as a cart-and-
pole system, which learned to stabilize within seconds. The
devised qualitative plan offered a comprehensive explana-
tion of the used control strategy. [7]
Discovering bipedal walking. A humanoid robot was
given a task to move forward a certain distance. As a means
to achieve this goal, the robot discovered a qualitative con-
cept of bipedal walking, which was successfully executed.
4 Conclusion
The thesis proposes a novel approach to qualitative plan-
ning and plan execution in continuous robotic domains.
The proposed method is domain independent and allows
qualitative models to be used without additional numerical
training. While the proposed qualitative approach might
not be feasible for all robotic problems, it has been success-
fully demonstrated with a variety of problems on different
types of robots. The results were published in several peer-
reviewed publications.
References
[1] Wiley T., Sammut C., Bratko I. (2016) A Planning
and Learning Hierarchy using Qualitative Reasoning
for the On-Line Acquisition of Robotic Behaviors,
Advances in Cognitive Systems 4, pp. 93–112.
[2] Šoberl, D. (2021) Automated planning with induced
qualitative models in dynamic robotic domains, PhD
thesis, University of Ljubljana. https://repozitorij.uni-
lj.si/IzpisGradiva.php?id=126285
[3] Kuipers B. (1986) Qualitative simulation, Artificial
Intelligence 29(3), pp. 289–338, DOI: 10.1016/0004-
3702(86)90073-1
[4] Šoberl D., Bratko I. (2017) Reactive Motion Plan-
ning with Qualitative Constraints, Advances in Arti-
ficial Intelligence: From Theory to Practice. IEA/AIE
2017. Lecture Notes in Computer Science, vol 10350,
Springer, Cham, pp. 41–50. DOI: 10.1007/978-3-
319-60042-0_5
[5] Šoberl D., Žabkar J., Bratko I. (2015) Qualitative
Planning of Object Pushing by a Robot, Foundations
of Intelligent Systems. ISMIS 2015. Lecture Notes in
Computer Science, vol 9384, Springer, Cham, pp.
410–419. DOI: 10.1007/978-3-319-25252-0_44
[6] Šoberl D., Žabkar J., Bratko I. (2020) Learning to
Control a Quadcopter Qualitatively, Journal of Intelli-
gent & Robotic Systems 100(3), pp. 1097–1110. DOI:
10.1007/s10846-020-01228-7
[7] Šoberl D., Bratko I. (2019) Learning Explainable
Control Strategies Demonstrated on the Pole-and-
Cart System, Advances and Trends in Artificial In-
telligence. From Theory to Practice. IEA/AIE 2019.
Lecture Notes in Computer Science, vol 11606,
Springer, Cham, pp. 483–494. DOI: 10.1007/978-3-
030-22999-3_42