Whereas the semantic properties of the approach
have been investigated in some detail, its algorithmic
and computational properties are still challenging. To
be useful in practice, the probabilistic DLs obtained
by applying this approach need to be equipped with
effective reasoning procedures. Thus, another main
emphasis of this project is on investigating computational properties (decidability and complexity) of the
probabilistic logics obtained by instantiating the
approach in particular with DLs of different expressive power.
We plan to apply these methods to set up and use
probabilistic ontologies for autonomous vehicles in
cellular transport systems and warehouse planning in
The principal investigators are Franz Baader and
Planning and Action Control for
Robots in Human Environments
The aim of this project is to develop approaches that
help a robot operate in an autonomous and flexible
manner in human-centered environments. One of
our goals is to assist human users with everyday tasks
in a way that aligns with their personal preferences,
for example, regarding how to organize objects or pri-
oritize tasks. We address this goal by exploring tech-
niques from recommender system theory to enable a
service robot to learn such preferences in a lifelong
manner and to tailor its behavior to its environment.
This approach allows us to make task (goal) recom-
mendations to high-level planners in order to trade
off user preferences with resources such as time or
This point brings us to another focus of this project, automated planning, where we address the challenge of enabling planners to efficiently react to
changing goals and situations. To that end, we adopt
an anytime planning approach in which we incrementally produce promising plan prefixes, start executing them, and in the meantime continue to refine
the plan in the background. We also investigate
heuristics that efficiently guide the search for solutions by abstracting away from geometric states and
planning at the symbolic level.
In the context of so-called continual planning,
symbolic and geometric reasoning are typically interleaved through semantic attachments. Here, we aim
to improve planning efficiency by extending geometric reasoners to incorporate relevant symbolic information. Additionally, we are developing approaches
that enable a robot to continuously update its knowledge by learning manipulation skills (for example,
how to grasp a new object) from nonexpert user
demonstrations, and encoding these skills both symbolically (for example, as plan operators) and in continuous terms.
Another approach to increase the efficiency of
planning considers the use of macro-actions. The idea
is that plans for typical household tasks such as clear-
ing a table often share short sequences of actions. We
identify such plans from a large collection of previ-
ously computed plans and turn the shared actions
into macro-actions so as to speed up future planning.
We have shown that this sorting can be done in a
provably correct way for an expressive fragment of
the Planning Domain Definition Language (PDDL).
In other work, we are endowing robots with the
ability to deal with task interruptions, for example,
briefly interrupting a tidying-up task to open the door
for a guest. Our continual planner handles this task
switching by enabling new searches in the changed
situation to reuse the explored space from previous
planning requests. Additionally, our high-level controller is able to store what is believed before the
interruption and to restore it afterwards to facilitate
resuming the stopped task. To support these and other features, we have developed a generic robot memory capable of remembering, updating, and forgetting
both short- and long-term information that ranges
from abstract symbolic representations to raw sensory data.
Finally, we are closely collaborating with our partners in the Advanced Solving Technology project in
the domain of multiagent production logistics as
modeled in the RoboCup Logistics League (RCLL).
This partner project involves addressing several challenges to improve cooperation within a fleet of robots
and hence optimize material flow in a dynamic factory environment. We aim to achieve this optimization using a centralized, global-scope planner based
on reactive ASP.
The principal investigators are Wolfram Burgard,
Gerhard Lakemeyer, and Bernhard Nebel.
We believe that the hybrid methods developed in the
Hybris Research Unit substantially increase the appli-
cability of symbolic reasoning methods to real-world
applications. Some of the principal investigators of
the project are currently working on a follow-up pro-
posal that will focus on applications in logistics. We
consider logistics an area where hybrid techniques are
particularly promising and beneficial. For more infor-
mation on the Research Unit and links to publica-
tions, visit www.hybrid-reasoning.org.
Gerhard Brewka is a professor and chair of Intelligent Systems at the University of Leipzig. He is an ECCAI fellow.
Gerhard Lakemeyer is a professor in the Department of
Computer Science and head of the Knowledge-Based Systems Group at RWTH Aachen University.