The work reported in this dissertation has promoted further research on automatic motion synthesis for computer animation and on locomotion learning for artificial life. Grzeszczuk and Terzopoulos [Terzopoulos, Tu and Grzeszczuk1994b, Grzeszczuk and Terzopoulos1995] have developed a learning technique that automatically synthesizes realistic locomotion for physics-based models of animals. This technique specifically addresses animals with highly flexible and muscular bodies, such as fish, rays, dolphins, and snakes. In particular, they have established an optimization-based, multi-level learning process on top of the motor system of the artificial fish. This process forms an additional ``locomotion learning'' center in the artificial fish's brain (see Fig. ). This center automatically learns effective motor controllers for the artificial fish biomechanical model, and abstracts them into suitably parameterized form. On the one hand, the learning center enhances the functionality of our animation system by subsuming the original laborious hand-crafting of motor controllers. On the other hand, equipped with the locomotion learning ability, the artificial fish has now `evolved' into a more complete artificial life form.
Figure: Locomotion learning center in the brain of the artificial fish.
The ability to learn also leads to the possibility of automatically generating simple sensorimotor tasks for artificial animals, such as the fish. First steps along these lines have already been made: Greszczuk and Terzopoulos Radek95 imbue the artificial fish with the ability to learn to maneuver and reach a visible target. This is done by enabling the animals to learn to put into practice the compact, efficient controllers that they have previously learned. In this way, Grzeszczuk and Terzopoulos have developed dolphin models that can learn to perform a variety of ``SeaWorld stunts''.
|Xiaoyuan Tu||January 1996|