The representation of the world in which an autonomous robot interacts with other objects within the world requires the integration of perception, reasoning, and action. The challenge of designing a suite of ontologies that are expressive enough to support perception and reasoning is a preeminent challenge for the knowledge representation and applied ontology communities. In this paper, we propose a series of challenge problems oriented around the ways in which a robot encounters the physical world. We outline an approach to the design and application of ontologies that can be used to represent the problems and their solution.
Imagine waking up in a well-lit, unrecognizable room. Although terrifying and stressinducing, a quick scan should allow you to quickly devise a plan of action. You are able to discern what sharp objects to avoid, what objects are movable, ”free space” and what would lead to a ”free fall”. Much of these common-sense reasoning is accumulated through trial-and-error over the years. However, even more of this common-sense is taught to us – for instance, we don’t need to get stabbed by a knife ourselves to understand that it is fatal. A small portion of our intelligence is genetic – for instance, we figure out how to walk and grab things without explicit teaching or extensive experience. This combination of spatial, temporal and physical knowledge allows us to reason about the properties of spaces and objects that we have never encountered. The question is, what is the minimum(albeit non-exhaustive) knowledge needed to represent a robot as an embodied cognitive agent, one that possesses a physical form, is aware of it, and interacts with other physical objects and other embodied cognitive agents? To this end, we will explore several scenarios and propose relevant ontologies.
Assuming a robot has successfully replicated our spatiotemporal and physical reasoning capability, what is the minimum we can expect it to be able to do? In this section, we present an initial set of reasoning problems that can be used to determine the suite of ontologies that we will need.
We believe that these problems address the minimum ontological commitments for a robot to replicate our ability to reason about space and movement. The next step is to design the weakest ontology possible for the robot to reason about solutions to the above set of problems.
Organizing knowledge around robotics is not a novel pursuit. Below we explore three largely known ontologies within the field of robotics.
CORA is an ontology specifically focused on the fields of industrial robotics, service
robotics, and autonomous robotics. It formally specifies mechanical components of
robots and robotics, types of robots and types of robot interactions. This is ideal for
classifying different types of robotic designs and movements. This is achieved through an
alignment with upper ontologies such as SUMO, and OWL. [
The OpenRobots Common Sense Ontology is closely related to KnowRob, both in
sharing of concepts and design concept. However, it differs from KnowRob in that it focuses
on human-robot interaction. It is aligned with the now defunct OpenCyc upper ontology.
It achieves its purpose by developing two main ontologiesn- one focusing on the theory
of the mind and one of ”common-sense”. The former helps develop context and make
sense of ambiguity in instructions, whereas the latter promotes an ability to reason about
actions to take. As such, ORO’s goal differs slightly from our proposed problem, as we
focus on problems in terms of robot-physical world interaction. [
Given the combined capability of these ontologies, one may wonder why we are proposing yet another ontology for robotics. However, it should be noted that in terms of application, we differ from ORO, in the sense that we focus on robot-physical world interactions as compared to robot-human interaction. Moreover, we differ from CORA in the sense that we do not seek to create a robotics ontology, but rather ontologies for robots (that will support common-sense reasoning). In addition, all of the above ontologies do not contain extensive first-order axiomatizations, which limits expressivity. KnowRob cicumvents this through their hybrid, software-oriented approach. Although effective, system complexity and different programming languages prove a barrier in reusability. As such, our design of a reasoning framework purely based on first-order logic is novel.
How can we be sure that our set of designed ontologies are right for the application?
Following [
The PSL Ontology [
Within the PSL Ontology, an occurrence tree G is a partially ordered set of activity occurrences, such that for a given set of activities, all discrete sequences of their occurrences are branches of the tree. An occurrence tree contains all occurrences of all atomic activities; activity occurrences that are elements of an occurrence tree are called arboreal activity occurrences.
Most applications of process ontologies are used to represent dynamic behaviour in the world so that intelligent agents may make predictions about the future and explanations about the past. In particular, these predictions and explanations are often concerned with the state of the world and how that state changes. The PSL core theory Tdisc state is intended to capture the basic intuitions about states and their relationship to activities.
Within the PSL Ontology, the notion of state is represented by reified fluents. Intuitively, a change in state is captured by fluents that are either achieved or falsified by an activity occurrence. The prior relation is used to specify the fluents that are intuitively true prior to an activity occurrence and the holds relation specifies the fluents that are intuitively true after an activity occurrence.
Furthermore, a fluent can only be changed by the occurrence of activities. Thus, if some fluent holds after an activity occurrence, but after an activity occurrence later along the branch it is false, then an activity must occur at some point between that changes the fluent. This also leads to the requirement that the fluents holding after an activity occurrence will be the same fluents that are prior to any successor occurrence, since there cannot be an activity occurring between them.
The key step in applying the PSL Ontology to the problems in Section 2 uses the
approach presented in [
Axioms in the ontology are mapped to state constraints, which are universal sentences containing only prior literals and a unique activity occurrence variable. For any domain ontology T , there exists a set of state constraints known as the domain state ontology. The domain process ontology is designed by focussing on the relationship be2This paper uses the axiomatization http://www.mel.nist.gov/psl/psl-ontology/. of the
PSL
Ontology found at tween models of the domain ontology and the states associated with activity occurrences in the occurrence tree.
We can now provide a formal specification of the reasoning problems from Section 2, and identify the ontologies that are required for the specification.
The first step in the formalization of Problem 1 is to identify the basic ontological commitments that are required to specify the problem.
We begin with the existence of a room that contains a set of objects. As observed
by [
The next aspect of Problem 1 focusses on the statement ”that the robot can take from
one location to another”. In other words, we need an ontology that can represent how
objects can change location, that is, an object can occupy one region at one point in time,
and then occupy a different location at another point in time. The design of an ontology
that satisfies this requirement is based on the methodology of [
(9o) holds(loc(Robot; R4); o) The Motion Domain Process Ontology classifies activities by the ways in which objects can change their location: (8a)motion(a) ((8o)occurrence o f (o; a) (9x; r1; r2) f alsi f ies(o; loc(x; r1)) ^achieves(o; loc(x; r2)) (5) The spatiotemporal objects that represent the path that the robot takes is axiomatized as a class of complex activities within the PSL Ontology: (8a) path(a) ((8a0) subactivity(a0; a) ^ primitive(a0) motion(a0)) (6) Finally, the constraint in Problem 1 is axiomatized by: (8x; y; o; o1; a) path(a)^occurrence o f (o; a)^subactivity occurrence(o1; o) :holds(contact(x; y); o) (7)
To identify which ontologies are needed to satisfy Problem 2, we can first attempt to break down the question as per Problem 1. Again, we begin with a room containing a set of objects, except with the addition of Article A. The next aspect of the problem then explains the role of Article A - the robot has to get across the room, without bumping any of the obstacles, while carrying Article A.
Clearly, this problem is an extension of Problem 1. In addition to understanding the
relation between an object and the physical space it occupies, the robot needs to now
understand the relation between itself and another object. Our approach to this problem
utilizes the mereotopology of connected objects proposed in [
A typical approach to the application of ontologies to the representation and solution of
commonsense reasoning problems is to use the strongest possible ontology (such as full
Euclidean geometry or ordered real closed fields) to represent the world in which a robot
is situated. In this paper we have taken a different perspective and sought the weakest
possible ontology that is sufficient to axiomatize the intended semantics presupposed by
the natural language statements of challenge problems for robotic knowledge
representation and reasoning. In particular, we have seen how mereotopology, location, and
process ontologies are sufficient to axiomatize the fundamental ontological commitments of
robot path planning and limited interactions of robots with physical objects in their
environment. In the spirit of the Physical Turing Test [