Morphological Computation in PACE

Traditionally, in IT applications in robotics, the focus has been on the control algorithm (or the neural system in biology). More recently there has been an increasing interest in the notion of embodiment which is about the relation between the body (morphology and materials), the brain (the control), and the environment. It turns out that part of the desired functionality can be "outsourced" to morphological and material aspects of a system, or the interaction with the environment. For example, the non-homogeneous arrangement of the facets in an insect eye perform a kind of "pre-processing" of the sensory stimulation which facilitates later processing by the brain. An example on the actuator side are the intrinsic material properties of the muscle-tendon system: if I twist my arm and I let it go, it will turn back into its natural position, but not so much because of neural control (although in humans, the brain is always at least to some extent involved), but due to the properties of the muscle-tendon system as a damped spring system. Again, the brain is outsourcing some of its computation (control) to the morphological and material properties of the limbs. In the passive dynamic walker, a robot that walks down a ramp without control and actuation, the "computation" is performed by the mechanical feedback generated as the walker interacts with the environment (an instance of morphological computation called self-stabilization) (Pfeifer et al., 2007). 


Another way of saying this would be that the computation is performed by the attractor dynamics of the mechanical system. Semi-permeable membranes are also instances of morphological computation. Another class of examples are modular robotic and self-assembling systems. Macroscopic modules (such as HYDRA or MTRAN modules) at the 10cm scale can be controlled by a microprocessor and actuators on the basis of a sensory-motor loop. When moving to substantially smaller scales, other technologies have to be used and morphological and material properties have be exploited to achieve the desired functionally, i.e. more of the computation has to be "outsourced" to morphology and materials because the relevant properties can no longer be directly controlled. Self-assembly is closely related to parallel process control.  


In PACE we investigated fundamental processes such as sorting of objects or establishing non-Fickian gradients in diffusion systems controlled only by morphology-dependent interactions between the system components, either with themselves or with the environment. Now, how can morphological computation be characterized in general terms? A precise definition is still lacking but two types can be distinguished: (i) where the problem is in digital form (e.g. finding a Hamiltonian path), the computation is performed by a physical substrate, and then there is a readout process which translates the problem back into a digital one. (ii) where the embodiment is directly exploited to contribute to the behavior system (e.g. robots capitalizing on morphology and materials, selective membranes, self-assembly), without there being a digital problem and finally a readout. Examples of the former is DNA or molecular computation, instances of the latter are the robotics and self-assembling systems. We talk about computation whenever three properties are clearly present in a system. First, we can sensibly talk about input and output. Second, we should be able to the notion of programmability. And third, there has to be a kind of "teleological embedding", i.e. the system has to have some kind of desired functionality. These ideas can be applied to all examples given above in a straightforward manner. 


PACE motivated to analyze the paradigm of morphological computation by contrasting two of its most relevant features with conventional computing:

  • In a conventional computation the relevant aspects of the physical reality underlying the process to control are mapped onto a digital representation. It is this representation, which is processed further by employing some sort of algorithm. The outcome of this algorithm is translated back into motor signals for the actuators. Control by conventional computation therefore implies a number of non-trivial requirements, which are not rooted in the process itself but related to its representation. First, the mappings between the physical system and representation have to be implemented. Second, the problem to be solved employing algorithms on the digital representation of the physical situation has to be efficiently tractable. We emphasize that tractability is often related to the properties of the representation and not the problem itself: It is much harder to calculate the dynamics of a non-linear system than a linear one, though for nature this distinction is irrelevant, both processes just take place following the laws of physics. It is a key objective of morphological computation to allow other than digital representations of a problem and therefore to exploit nature's indifference towards problems of traditional numerical analysis.

  • The number of variables employed is of high relevance for a conventional computation. Nature in contrast is inherently parallel. This does not only mean that e.g. many particle processes pose no speed or communication problem as it is the case in conventional computation, but also that soft objects, characterized by more than the six parameters of a rigid body, can be treated without loss of efficiency. Also the interaction of stiff objects with non-trivial shape, a (digital) computationally intensive task, is readily done in the setting of morphological computation.


A formal definition of the term "morphological computation" is presently lacking. This at least partially, because the term "representations other than digital" is generically an open concept. Subclasses of morphological computation, such as shape-based computation, may be defined by giving complete descriptions of the type of problem representations they are able to process. It is emphasized that the representation issue covers only one side of the concept of morphological computation. The other is given by the types of algorithms that can be performed in an embodied situation. Not all conceivable algorithms can be implemented in the physical world, but those which can take profit of the inherent parallellism of physical laws. This consideration also points towards the difference between a (metaphorically) "hardwired" and a "programmable" morphological computation. "Pre-processing" information by sensors exploiting material properties is in our interpretation a morphological computation, because it are the sensory stimuli as a representation of the physical processes in the environment that are processed. However, such a pre-processing is not necessarily programmable. In order to achieve programmability, we need a kind of compiler that is able to translate certain classes of problems into morphological settings. In any case, such a compiler requires a thorough understanding of the possible morphological algorithms and the way they can be combined. In the course of PACE, considerable progress was made with respect to morphological algorithms (sorting) as well as the realization of the hardware for vesicle-based chemical process management.

© 2004-2008 All rights reserved by PACE Consortium .   Email.   Web Managers: U. Tangen & J. S. McCaskill