It is perhaps human nature to think in terms of contrasting dyads: good and evil, left and right, or ying and yang. Cognitive neuroscience has certainly not escaped this kind of dualistic thinking and over the years has generated a number of dyadic concepts such as explicit versus implicit knowledge, declarative versus procedural memory and, of course, perception versus action. Although one has to be careful not to oversimplify things, such distinctions can sometimes serve a useful function in emphasizing important differences in the way in which processing might unfold. The distinction between vision-for-perception and vision-for-action is a good example of this. By examining the differences in the way in which visual information is transformed for perceptual representation, on the one hand, and the control of action, on the other, Goodale and Milner (Reference Goodale and Milner1992) were able to make sense of an otherwise confusing and apparently contradictory set of observations in human neuropsychology and monkey neurophysiology. In the target article, Dijkerman & de Haan (D&dH) have used this same dialectic method and applied it to the somatosensory system, synthesizing a clear and compelling story from a literature that at first glance seems equally disparate and unconnected. Indeed, the parallels drawn between the organization of the somatosensory system and the visual system are remarkably close, reflecting no doubt the fundamental differences in the nature of the demands that perception and action put on any sensory system.

D&dH provide plenty of evidence that somatosensory information is processed differently for perception than it is for action – and make explicit the idea that this division of labour is similar to that seen in the visual system. But they go on to suggest that the separation in the pathways (and the transformations they perform) may be “less distinct” in the somatosensory system than in the visual system. Although it is true that the visual streams appear to be more anatomically distinct than the somatosensory streams, the dorsal and ventral visual pathways are likely to be just as interactive as those in the somatosensory system. Indeed, in their 1995 monograph, Milner and Goodale made it clear that the two systems were not “hermetically sealed” from each other, but were closely coupled in the production of adaptive behaviour (Milner & Goodale Reference Milner and Goodale1995). Certainly, D&dH acknowledge that such interactions occur, but perhaps it is worth emphasizing that the integration between processing in the two visual streams is more intimate than widely believed. A case in point is the way in which high-level visual processing in the ventral stream plays a fundamental role in certain aspects of action control. Moreover, as we shall see, this processing also depends on integrating visual and somatosensory information.

When we reach out to pick up an object, we need to open and orient our hand and fingers appropriately in flight – and when we make contact with the object, we need to apply just the right amount of force so that we do not damage the object and, at the same time, it does not slip out of our hand. The computations underlying the former are largely mediated by mechanisms in the dorsal stream, since information about the size and overall shape of the object, and its disposition with respect to our hand, is readily available from the retinal array (Milner & Goodale Reference Milner and Goodale2006). But any calculations about the initial grip and load forces that need to be applied to the object must also reflect the object's mass, density, surface friction, and compliance – characteristics that cannot be derived from the retina directly. For this reason, it is highly unlikely that the dorsal stream by itself could compute this information, and instead the specification of the required forces would have to rely on stored information about the material properties of the particular goal object (or, in the case of novel objects, on predictions derived from stored information about similar objects). The retrieval of this stored information would depend on visual processing in the ventral stream, which would enable us to recognize a goal object and/or its material properties on the basis of a broad range of cues, including both shape and surface properties. Thus, real-time visuomotor control would appear to depend on interactive processing in the dorsal and the ventral streams – in a manner that is not dissimilar to what D&dH have suggested happens in the two functional networks of the somatosensory system.

It is worth emphasizing, too, that the control of the forces required to pick up an object is heavily dependent on the close integration of visual and somatosensory processing. Although visual inspection of an object enables us to program the amount of grip and load forces needed to lift the object, once we have the object in our hand, somatosensory feedback from our fingers enables us to fine-tune the forces, which were initially based on visual processing alone. The result is force output that is more closely matched to the intrinsic properties of the object. Indeed, internal forward models of this kind have been influential in demonstrating the interactions between visual and somatosensory processing in the control of skilled movements (Flanagan & Johansson Reference Flanagan, Johansson and Ramachandran2002). A key aspect of these forward models is the initial prediction that the model makes, which is used to program relevant behaviour. In the case of the forces required to grasp and lift an object, this prediction is based on stored information about the material properties of that object derived from the somatosensory system during earlier encounters with the object (or objects made of the same material).

But how does one access this stored knowledge about material properties? One critical route that has been overlooked, at least in the object-recognition literature, is via the visual surface properties of the object (colour, texture, and specularities, for example) that are closely correlated with the object's material properties. In a recent functional brain imaging study, Cant and Goodale (Reference Cant and Goodale2007) showed that there are distinct pathways within the ventral stream that appear to be specialized for dealing with the surface properties of an object as distinct from its form. Whereas form is processed more laterally in the lateral occipital area (area LO), surface properties (particularly texture) are processed more medially in the collateral sulcus and the lingual and parahippocampal gyri. Presumably, the identification of the surface properties of an object, particularly in the case of novel objects, allows us to calibrate the initial forces required to lift the object by invoking an association between those surface properties and stored information derived from the somatosensory system about previous encounters with the material from which that object is made.

Thus, not only do the ventral and dorsal streams work closely together in the control of skilled movements, but they also have intimate associations with the somatosensory system in both programming and executing this control. Nevertheless, it is important to emphasize that the dorsal stream exercises its control over actions in a direct and largely bottom-up fashion, whereas the ventral stream makes its contributions to this control via indirect and top-down mechanisms. Presumably, such distinctions in the nature of the control will also be evident within the somatosensory system. In any case, the division of labour that D&dH have put forward for the somatosensory system is an important first step in understanding how somatosensation-for-perception and somatosensation-for-action work together in the production of adaptive behaviour.