What embodied cognition is not

Traditional theories in psychology place all the responsibility for generating our behaviour in the brain; perception is the input to a computational, representational system that mentally transforms the input into motor commands. Many researchers treat embodied cognition as the idea that the contents of these mental states/representations can be influenced by the states of our bodies.

For example, Eerland, Guadalupe & Zwann (2011) recently suggested that leaning to the left makes you underestimate the magnitude of things, because this state of the body biases you towards the left hand end of the mental number line they claim we use to generate estimates of magnitude (more here). You can also demonstrate the connection between 'mind' and 'body' goes the other way; Miles, Lind & Macrae (2010) showed that thinking about the future made you sway forwards and thinking about the past made you sway backwards; they suggest this sway is caused by the underlying metaphor of 'past is behind us' and 'the future is ahead of us' (more here).

There are numerous problems with these studies; the primary problem, however, is that in both cases the mental content is assumed to the same as it would be if you were doing non-embodied cognition. This research remains business as usual, with a couple of embodied bells and whistles—all the hard work of generating behaviour is done in the brain, it's just that this work can be biased by what the body is up to. However, the hypothesis that the body can play a role in cognition is actually much more radical than this. By expanding the resources available to solve a task from simply the brain to include the body (with its perceptual and motor systems), we've opened up the possibility that we can solve a task in a very different way than a brain by itself might solve the task.

 

What embodied cognition is

The kind of embodied cognition we advocate (more here) is the claim that the brain, while important, is not the only resource we have available to us to generate behaviour. Instead, the form of our behaviour emerges from the real-time interaction between a nervous system in a body with particular capabilities and an environment that offers opportunities for behaviour and information about those opportunities. The reason this is quite a radical claim is that it changes the job description for the brain; instead of having to represent knowledge about the world and using that knowledge to simply output commands, the brain is now a part of a broader system that critically involves perception and action as well. The actual solution an organism comes up with for a given task includes all these elements.

Our favourite example of this right now is 'the outfielder problem' (more here). This is the question of how a baseball outfielder can catch a fly ball—how do they manage to get to the right place at the right time? The disembodied, computational solution notes that in order to predict where the ball will land, you need a model of the projectile motion of the ball and some information about its initial conditions as it came off the bat (speed, direction, etc). Perception provides this input, the brain uses a representation that implements the model to predict the landing location and then commands the body to move the right location.

The embodied solution instead begins with a task analysis; what resources are available to a mobile, perceiving-acting organism to solve the task? Do we need to compute anything, or can we simply directly use perceptual information to guide our interception? It turns out that careful consideration of the task reveals two potential informational strategies that are consequences of the parabolic arc the fly ball takes through the air.

The first works if you are in a direct line with the arc of the fly ball; if you run so as to make the (actually accelerating) ball appear to move at constant velocity, it turns out you will end up in the right place at the right time. If you are off the direct line, you can move so as to make the (actually curved) trajectory of the ball appear linear. There is no prediction or internal model required; instead, the outfielder solves the prediction problem by moving in a particular way.

Evidence strongly supports this latter prospective control strategy—people move in the kinds of curved paths it predicts (McBeath, Shaffer & Kaiser, 1995), as do dogs (Shaffer, Krauchnas, Eddy & McBeath, 2004). In effect, the work done by a computation in the disembodied account is instead done by the way you move in the embodied account; embodiment does work to solve problems that was typically assumed to be all done in the head.

Our other favourite way to describe the difference between the traditional, non-embodied computational approach and embodied cognition is to point to two very different kinds of robots (more here). Robots have an advantage in that we know exactly what went into making them—we know if they are computing sophisticated solutions, or simply moving under the control of simple rules. We can then look to see what their behaviour is like and relate that to biological organisms like ourselves.

Honda's ASIMO literally implements a traditional cognitive, computational approach. Everything it does is the output of complex internal programmes which control everything he does. Honda are fond of trotting him out to dance, run, and climb stairs; he can do all this, but it's very fragile. Minor disruptions throw him entirely (e.g. a minor error in foot placement and he falls (http://www.youtube.com/watch?v=ASoCJTYgYB0); hide his pre-set landmarks with a little clutter and he completely fails to navigate his way across a room). He's slow, and inefficient; if you knock him, he needs time to recompute his behaviour or else he falls, and he often doesn't have the time.