Current Research
Physical simulation is quickly becoming de rigeur in interactive simulations ranging from high-budget computer games to “serious games” that educate students, train soldiers and rehabilitate the injured. Nevertheless, systems of control for dynamic human characters in these simulations are still in their infancy. Humans present exceptional difficulties to control both because they are biped and thus not statically stable during most motions, and because the standard for fluidity and fidelity is high-- virtual humans should move like natural humans. This is a surprisingly difficult problem that has not been adequately solved in either robotics or computer animation.
I'm interested in combining low-level techniques adapted from robotics with machine learning techniques to allow the automatic synthesis of controllers. The low-level techniques are based on the efficient computation of composite inertia tensors, applied-torque compensation, and several metrics indicative of balance (e.g., ZMP, ZRAM). The higher-level is based on the artificial evolution of neural networks. Novel representations and abstractions derived from the lower-level control techniques reduces and smoothes the problem search-space. Early results have provided initial validation of the neuroevolutionary approach for controlling bipeds for initiating locomotion from standing, walking short distances, and balancing against small random perturbations.
Current work examines the possibility of generating objective functions automatically by classifying system state-space according to known control strategies. Areas where no existing strategy is known inform the automatic creation of new objective functions.