Experiments in animation control by neural networks

The use of neural nets with an associated learning paradigm was investigated for the purposes of animation control. Neural control schemes were set up and used to control simple skeletal figures. In most cases, the methods eventually converged to a controller that performed very well, even when the learning algorithm became trapped at local minima. However, there is a limit to the learning rate, or else instability arises. Nevertheless, a useful model of the inverse dynamics of the controlled object can be captured fairly quickly, but fine adjustments take some time. Most skeletal figures one might want to control combine kinematic redundancy, inherent instability and limited controllability, and the unicycles studied had all of these properties. This meant that a new neural control scheme had to be devised. The inputs of the neural net were taken from both the reference trajectory and the status of the object being controlled, whilst initially only attempting to control part of the object. Additional experiments were carried out successfully with a weak conventional controller that simply let the unicycle fall over, producing a natural learning mode in which the unicycle begins by falling over and gradually gains the skill required to stay upright and control its motion as training goes on. Conventional control schemes in Cartesian space were also considered. After some experimentation with direct user control of the torques and the reference input, an intuitive control scheme that combined the neural controller and a trajectory generator was defined. Quite realistic and elegant motion was eventually achieved.