Previous: Symbol grounding: A non-Tarskian semantics
Up: The GLAIR Architecture
Next: Alignment
In section we already mentioned the use of embodied
representations at the Perceptuo-Motor level. We now look at the principle
of embodiment from a more abstract point of view.
One of the most general motivations behind our work is the desire to be
able to ``program'' a robotic autonomous agent by requesting it to do
something and have it ``understand'', rather than telling it how to do
something in terms of primitive motions with little or no
``understanding''. For instance, we want to tell it to go find a red pen,
pick it up, and bring it to us, and not have to program it at a low level
to do these things. One might
say that we want to communicate with the robot at the speech act
level. To do this, the agent needs a set of general-purpose perceptual and
motor capabilities along with an ``understanding'' of these capabilities.
The agent also needs a set of concepts which are similar enough to ours to
enable easy communication. The best way to accomplish this is to endow the
agent with embodied concepts, grounded in perception and action.
We define embodiment as the notion that the representation and extension of high level concepts is in part determined by the physiology (the bodily functions) of an agent, and in part by the interaction of the agent with the world. For instance, the extension of color concepts is in part determined by the physiology of our color perception mechanism, and in part by the visual stimuli we look at. The result is the establishment of a mapping between color concepts and certain properties of both the color perception mechanism and objects in the world. Another example is the extension of concepts of action: it is partly determined by the physiology of the agent's motor mechanisms, and partly by the interaction with objects in the world. The result is the establishment of a mapping between concepts of action and certain properties of both the motor mechanisms and objects in the world (what we might call ``the shapes of acts'').
At an abstract level, the way to provide an autonomous agent with human-like embodied concepts is to intersect the set of human physiological capabilities with the set of the agent's potential physiological capabilities, and endow the agent with what is in this intersection. To determine an agent's potential physiological capabilities, we consider it to be made up of a set of primitive actuators and sensors, combined with a general purpose computational mechanism. The physical limitations of the sensors, actuators, and computational mechanism bound the set of potential capabilities. For instance with respect to color perception, if the agent uses a CCD color camera (whose spectral sensitivity is usually wider than that of the human eye), combined with a powerful computational mechanism, we consider its potential capabilities wider than the human ones, and thus restrict the implemented capabilities to the human ones. We endow the agent with a color perception mechanism whose functional properties reflect the physiology of human color perception. That results in color concepts that are similar to human color concepts. With respect to the manipulation of objects, most robot manipulators are inferior to human arms and hands, hence we restrict the implemented capabilities to the ones that are allowed by the robot's physiology. The robot's motor mechanism then reflects the properties of its own physiology, rather than those of the human physiology. This results in a set of motor concepts that is a subset of the human one. Embodiment also calls for body-centered and body-measured representations, relative to the agent's own physiology. We provide more details on embodiment in GLAIR in [Hexmoor et al. 1993c].