This is an essay about the epistemology of biorobotics, which consists in making robotic models of animals with the goal of understanding biology (as opposed to solving robotics problems). It is a mine of references about the epistemology of modelling, and in particular what we mean by « model » and what is a « good model ».

One important point made in the essay is about the notion of abstraction. Robotic models, or computer models, are necessarily abstract in various ways (or more accurately, the connection with biology is at a quite abstract level ; the robot itself is not abstract). But abstraction is not the fundamental problem ; the problem is rather if things are simplified the wrong way. For example, theories of visual perception have been built based on presenting visual gratings to the animal and recordings neural responses, or asking a monkey to discriminate between two moving patterns of random dots. This is a massive abstraction of what visual behavior is. Which abstraction is more justified, simplifying the properties of neurons or getting rid of the ecological context entirely  ?

In my opinion, the most important point made in the paper about robotic models is that « making something that actually works [...] creates much of the hypothesis testing power of robotic models of biological systems. ». That is, a very important constraint that is included in such models, which is not included in most neural models, is that it has to be able to produce some kind of successful behavior. It is striking to see that most neural models of perception deal with trivial tasks (discrimination) or even no task at all (reproducing some experimental recordings), or even no sensory input at all (the « input » is a stimulus parameter). When neural models are compared to behavioral experiments (most precisely, psychophysical measurements, which do not correspond to any normal animal behavior), this comparison is done by the intermediate of a « decoder » or « ideal observer », abstract constructs which are not biological models. How can a model be biologically relevant if it does not explain behavior ? As Webb puts it, « the context of the behaviour of the organism is always included in a robot model, counteracting the tendency in biological studies to lose sight of this context in close study of small parts of the underlying mechanisms » and « robotic implementation specifically supports the consideration and integration of different levels of explanation because of its emphasis on requiring a complete, behaving system as the outcome of the model. ».

Another question is the interest of using actual robots rather than simulated organisms. The point, also made by Rodney Brooks, is that real environments are more complex than simulated ones, in ways that may not be anticipated: « even two-wheeled motor control has to cope with friction, bumps, gravity, and so on; whereas a six-legged computer simulation may restrict itself to representing only the kinematic problems of limb control and ignore the dynamics entirely. ». Adapted behavior in the real world is a more stringent test than in simplified environments.

Finally, this essay comes with 32 critical comments from various authors, which have different perspectives on the subject (it comes from a journal with open peer commentary).