Tutorial on Embodiment
5.3.2. Why is that cognition?*
As discussed by Clark & Grush (1999), forward models are the simplest instances of circuitry that emulates the world outside and thus stands for something that is not currently present in the sensory and motor states. Thus, we may want to attribute representation to such circuitry. A ‘decoupled' forward model that is not just a few steps ahead of the sensory-motor reality but that can be executed independently, in the brain only, can then be viewed as emulation/simulation of the interaction with the world, or world model. Interestingly, such a forward model can also be exploited to exercise embodied categorization, which we have presented in the previous section, in simulation. In other words, if the agent can predict the sensory consequences of its actions, it can also ‘imagine' circling around a big or small cylinder or catching a circle or diamond. The outcome of such internal simulation can be used to derive a perceptual judgment that would otherwise not have been possible. This is demonstrated by the agent of H. Hoffmann (2007) which uses such a ‘mental' rehearsal of driving in its environment to discriminate passages and dead ends.
Let us now wrap up the nature of representations and cognition that we are acquiring. Rather than representing static features (such as objects), dynamic interaction patterns, which involve the robot acting in the environment, are represented. Such representations are best viewed as motor-based. They are action-oriented, originate in the sensory-motor apparatus and remain intimately related with it (Clark & Grush, 1999; Pezzulo, 2007) - as opposed to symbolic AI representations that are world-centered. Whether we want to call these phenomena ‘cognitive' depends on our definition of cognition. Some views reject the cognitive/non-cognitive divide altogether, some include into the cognitive realm all kinds of adaptively valuable organism/ environment coupling (e.g., Thelen & Smith, 1994). While we consider these views equally legitimate, the view proposed by Clark & Grush (1999), among others, is that cognizers must display the capacity for environmentally decoupled thought and contemplation of options. This is exactly what a decoupled forward model provides: simulation of the world, or ‘mental imagery'. This phenomenon is believed to be at the core of grounded cognition (Barsalou, 2008; Gallese & Lakoff, 2005).
*This section has been adapted from Hoffmann & Pfeifer, 2011.
Barsalou, L. (2008), 'Grounded cognition', Annual Review of Psychology 59, 617-645.
Clark, A. & Grush, R. (1999), 'Towards Cognitive Robotics', Adaptive Behaviour 7(1), 5-16.
Gallese, V. & Lakoff, G. (2005), 'The brain's concepts: The role of the sensory-motor system in conceptual knowledge', Cognitive Neuropsychology 21, 455-479.
Hoffmann, H. (2007), 'Perception through visuomotor anticipation in a mobile robot', Neural Networks 20, 22-33.
Hoffmann, M. & Pfeifer, R. (2011), The implications of embodiment for behavior and cognition: animal and robotic case studies, in W. Tschacher & C. Bergomi, ed., 'The Implications of Embodiment: Cognition and Communication', Exeter: Imprint Academic, pp. 31-58.
Pezzulo, G. (2007). Anticipation and Future-Oriented Capabilities in Natural and Artificial Cognition. In M. Lungarella, F. Iida, J. C. Bongard, & R. Pfeifer, (eds.), 50 Years of AI, Festschrift (pp. 258-71). Berlin: Springer.
Thelen, E. & Smith, L. (1994), A Dynamic systems approach to the development of cognition and action, MIT Press.