v ision can be viewed as a passive, observational activity, or as one intimate ly related to action (for example, manipulation, navigation). In passive vision systems the camera providing the image input is immobile. In active vision systems observer-controlled input sensors are used.' Active vision results in much simpler and more robust vision algorithms, as outlined in Table 1.Another dimension for classifying computer vision approaches is reconstructive versus animate. In the reconstructionist or general-purpose paradigm, the vision task is to reconstruct physical scene parameters from image input, to segment the image into meaningful parts, and ultimately to describe the visual input in such a way that higher level systems can act on the descriptions to accomplish general tasks. During the last decade, substantial progress in reconstructionist vision has been made using both passive and active systems that exploit physical and geometric constraints inherent in the imaging process. However, reconstructionist vision appears to be nearing its limits without reaching its goal.An alternative to reconstructionist vision derives from the observation that biological systems do not, in general, perform goal-free, consequence-free viion.^ This observation suggests that vision may, of necessity, be a more interactive, dynamic, and task-oriented process than is assumed in the reconstructionist approach. Animate vision researchers, inspired by successful biological systems, seek to develop practical, deployable vision systems by discovering and exploiting principles that link perception and action. Animate systems use active vision and are structured as vertically integrated skills or behaviors, rather than as visual modules that try to reconstruct different aspects of the physical world.Despite the computational simplifications of the animate vision paradigm, a parallel implementation is necessary to achieve the required performance. Fortunately, many of the tasks in an animate vision system are inherently parallel. Inputs from multiple sensors can be processed in parallel. Low-level-vision algorithms are intensely data parallel. Planning and strategy algorithms frequently search a large state space, which can be decomposed into smaller spaces that are 0018-9162/92/0200-0012$03.00 0 1992 IEEE COMPUTER