Möller (1996,1999) suggested that the perception of shape (of an object) and space (surrounding an agent) is based on the simulation of actions and their sensory consequences. In this work, the one-way sensory processing is criticized to be unable to select behavior-relevant image structures, and further, to require an internal observer that needs to act upon a sensory representation to choose the appropriate behavior. As an alternative, perception is understood as anticipation. First, an agent learns the sensory consequences of its motor commands. Then, it simulates covert motor commands to obtain the behavioral meaning of sensory information.
In a behavioral task, a series of actions is chosen due to an evaluation of the predicted sensory state. Based on this idea, Gross et al. (1999) let a mobile robot navigate collision-free through a maze. In a recognition task, just the outcome of a simulated action sequence is analyzed. For example, a mobile agent would recognize a situation as a dead-end by mentally simulating movements and by predicting that no movements are possible beyond the dead-end. Here, perception is not based on matching visual cues of the dead-end to a prototype stored in the brain. Thus, the approach has the potential to generalize over dead-ends of different appearance and to solve the problem of object constancy (that is, an object can be recognized independently of the perspective). Further, objects can be classified based on the outcome of the simulation.
Such an approach requires the brain to be able to simulate actions without causing movements, to simulate sensation without receiving sensory input, and to associate action with the resulting changes in sensation. Hesslow (2002) reviewed evidence for all three requirements1.6. He even suggested that internal simulation is a ``mechanism for generating the inner world that we associate with consciousness'' (p. 246).
The ideas of Grush (2004) go in the same direction. In addition, he claimed that the simulation of covert motor commands is not enough because the outcome of a motor command depends on the current proprioceptive state. Thus, a covert motor command should act on an `emulator' of the body, which updates the simulated proprioceptive state. Such an emulator would, for example, explain the occurrence of a paralyzed phantom limb following pre-amputation paralysis: before the amputation, the emulator learned that any motor plan is mapped onto a constant proprioceptive state.
Consistent with all these simulation approaches are psychological experiments that show that action planning and perception (or sensory prediction) interfere with each other (Wexler and Klam, 2001; Wohlschläger, 2001; Prinz, 1997). Thus, action planning and perception seem to share a `common representational domain' (Prinz, 1997). Such a common representation might lie in the mechanism of sensorimotor simulation.