Cognitive models typically fail to interact with the external environment when performing a task. Where models have incorporated interaction this has not generally been implemented as a psychologically plausible mechanism. In the real world, where many tasks are now computer based, task context influences performance in all but the most trivial of tasks. Visual perception and motor action are therefore central to task performance, and hence to the generation of erroneous actions. The failure to take account of the constraints imposed by having to interact with the environment helps to explain why cognitive models perform tasks (and learn) too quickly.
We identify here a set of requirements to constrain the design of simulations of visual perception and motor action that form extensions to a cognitive architecture. These simulations operate at a symbolic level of abstraction such that the model can internally represent each object in the environment using a symbol. We also include a set of facilities to support debugging of the interaction mechanism, based on our experiences from earlier attempts at implementing such a capability.