Proceedings of the 2007 ACM symposium on Virtual reality software and technology
| Hi-index | 0.00 |
Motion graphs have gained popularity in recent years as a means for re-using motion capture data by connecting previously unrelated segments of recorded motion. The techniques for controlling character movement via motion graphs have largely focused on path planning which is difficult because of the dense connections found in the graph. In addition, current use of learning controller directly on motion graph are not meaningful due to intractably large learning space generated by the dense graph. We introduce a novel alternative which allows high level control of character behavior using a dual representation we called a state-annotated motion graph. This special motion graph is generated from labelled data and then bound to a finite state machine with similar labels. At run-time, character behavior is simply controlled by switching states. We show that it is possible to generate rich, controllable motion without the need for deep planning. And learning from the state-level is capable of producing new high-level behaviors with a week of unsupervised training. For our results, we demonstrate that simple state-switching controllers can be coded intuitively to create various effects applied to an interactive fighting testbed.