ABSTRACT
Numerical Solution of a Hamilton-Jacobi-Bellman Equation and its Application to Animation.
Gideon Amos, UCL
The trouble with computer animation of humans and other animals, is that
realistic motion and interactivity rarely go together. We are developing a
method that may go part way towards solving this problem. Starting with a
dynamic model of a simple 'toy' creature, the method generates a feedback
controller for stationary or repetitive behaviours. Together, the dynamic
model and controller can generate globally optimal trajectories at
interactive rates. The method solves a first order partial differential
equation, a Hamilton-Jacobi-Bellman equation, the solution of which exists
only in a specific sense. It is the continuous state equivalent of a dynamic
programming solution and is also a form of reinforcement learning. The
overwhelming limitation of this method is that the cost of the solution
scales exponentially with the number of degrees of freedom of the dynamic
model. A kd-tree based adaptive meshing stragegy is employed to improve the
scalability. We hope to demonstrate that this method can produce realistic
and engaging interactive animations.
Computer Science Home Page
Maintained by rbennett@cs.ucl.ac.uk