Learning the Affordances of Tools Using a
Behavior-Grounded Approach
Alexander Stoytchev
Department of Electrical and Computer Engineering
Iowa State University, Ames IA 50011, USA
alexs@iastate.edu
Abstract. This paper introduces a behavior-grounded approach to rep-
resenting and learning the affordances of tools by a robot. The affordance
representation is learned during a behavioral babbling stage in which the
robot randomly chooses different exploratory behaviors, applies them to
the tool, and observes their effects on environmental objects. As a re-
sult of this exploratory procedure, the tool representation is grounded in
the behavioral and perceptual repertoire of the robot. Furthermore, the
representation is autonomously testable and verifiable by the robot as
it is expressed in concrete terms (i.e., behaviors) that are directly avail-
able to the robot’s controller. The tool representation described here can
also be used to solve tool-using tasks by dynamically sequencing the ex-
ploratory behaviors which were used to explore the tool based on their
expected outcomes. The quality of the learned representation was tested
on extension-of-reach tasks with rigid tools.
1 Introduction
The ability to use tools is one of the hallmarks of intelligence. Tool use is fun-
damental to human life and has been for at least the last two million years.
We use tools to extend our reach, to amplify our physical strength, to transfer
objects and liquids, and to achieve many other everyday tasks. A large number
of animals have also been observed to use tools [1]. Some birds, for example, use
twigs or cactus pines to probe for larvae in crevices which they cannot reach with
their beaks. Sea otters use stones to open hard-shelled mussels. Chimpanzees use
stones to crack nuts open and sticks to reach food, dig holes, or attack predators.
Orangutans fish for termites with twigs and grass blades. Horses and elephants
use sticks to scratch their bodies. These examples suggest that the ability to
use tools is an adaptation mechanism used by many organisms to overcome the
limitations imposed on them by their anatomy.
Despite the widespread use of tools in the animal world, however, studies
of autonomous robotic tool use are still rare. There are industrial robots that
use tools for tasks such as welding, cutting, and painting, but these operations
are carefully scripted by a human programmer. Robot hardware capabilities,
however, continue to increase at a remarkable rate. Humanoid robots such as
Honda’s Asimo, Sony’s Qrio, and NASA’s Robonaut feature motor capabilities

similar to those of humans. In the near future similar robots will be working
side by side with humans in homes, offices, hospitals, and in outer space. It is
difficult to imagine how these robots that will look like us, act like us, and live in
the same physical environment like us, will be very useful if they are not capable
of something so innate to human culture as the ability to use tools. Because of
their humanoid “anatomy” these robots undoubtedly will have to use external
objects in a variety of tasks, for instance, to improve their reach or to increase
their physical strength. These important problems, however, have not been well
addressed by the robotics community.
Another motivation for studying robot tool behaviors is the hope that robotics
can play a major role in answering some of the fundamental questions about
tool-using abilities of animals and humans. After ninety years of tool-using ex-
periments with animals (see next section) there is still no comprehensive theory
attempting to explain the origins, development, and learning of tool behaviors
in living organisms.
Progress along these two lines of research, however, is unlikely without ini-
tial experimental work which can be used as the foundation for a computational
theory of tool use. Therefore, the purpose of this paper is to empirically eval-
uate one specific way of representing and learning the functional properties or
affordances [2] of tools.
The tool representation described here uses a behavior-based approach [3]
to ground the tool affordances in the existing behavioral repertoire of the robot.
The representation is learned during a behavioral babbling stage in which the
robot randomly chooses different exploratory behaviors, applies them to the tool,
and observes their effects on environmental objects. The quality of the learned
representation is tested on extension-of-reach tool tasks. The experiments were
conducted using a mobile robot manipulator. As far as we know, this is one of
the first studies of this kind in the Robotics and AI literature.
2 Related Work
2.1 Affordances and Exploratory Behaviors
A simple object like a stick can be used in numerous tasks that are quite different
from one another. For example, a stick can be used to strike, poke, prop, scratch,
pry, dig, etc. It is still a mystery how animals and humans learn these affordances
[2] and what are the cognitive structures used to represent them.
James Gibson defined affordances as “perceptual invariants” that are directly
perceived by an organism and enable it to perform tasks [2]. Gibson is not
specific about the way in which affordances are learned but he suggests that
some affordances are learned in infancy when the child experiments with objects.
For example, an object affords throwing if it can be grasped and moved away
from one’s body with a swift action of the hand and then letting it go. The
perceptual invariant in this case is the shrinking of the visual angle of the object