Adaptive behavior is supported by perception of the environment as much as it is by action. It is therefore crucial to model perception in artificial autonomous agents, including animated animals, humans and physical robots. Reynolds' ``boids'' maintained flock formations through simple perception of other nearby actors [Reynolds1987] while Matarić has demonstrated similar flocking behaviors with physical robots [Mataric1994]. The roach actor described by McKenna et al. Mckenna90 retreated when it sensed danger from a virtual hand. Renault et al. Renault90 advocate a more extensive form of synthetic vision for behavioral actors, including the automatic computation of internal spatial maps of the world. The virtual humans in Thalmann's work D-Thalmann95 have simulated simple visual, tactile and auditory sensors that enable them to perform tasks such as following a leader or greeting each other and even playing ball games.
Perception modeling for animation, in general, is very different from that for robotics. In an animation system, the detailed geometry of each scene can always be obtained by interrogating the virtual world model, without extensive sensory information processing. In a robotics system, however, this is not true. In fact, in order for a robot to obtain useful perceptual information, a visual process needs to be synthesized. This will include algorithms to infer 3D geometry from images, to identify shapes and to produce appropriate representations of objects, etc. On the other hand, in a typical animation system, since a database of all graphical objects exists and is accessible, the main purpose of modeling perception in animated figures is to enforce behavioral realism. This often only requires that the perceptual capability of the animal be modeled, such as the field of view and occlusion. Reynolds' model of boids (which only consists of simple modeling of limited field of view) is a good example.
Our artificial fishes are currently able to sense their world through simulated visual perception within a deliberately limited field of view. Subject to the natural limitations of occlusion, they can sense lighting patterns, determine distances to objects, and identify objects by inquiring the world model database. They are also equipped with secondary nonvisual modalities, such as the ability to sense the local virtual water temperature. More importantly, unlike previous perception models for animation, the artificial fish's perception induces an attention mechanism. This mechanism allows the fish to train its sensors in a task-specific way as well as to provide other important information for producing convincing behavior. Our model of perception is proven to be effective in generating realistic behaviors of the artificial fish.
|Xiaoyuan Tu||January 1996|