A neural network learns to generate realistic character animations in real-time by adding timing information to motion synthesis. The system handles walking, running, jumping, and climbing across varied terrain with minimal memory, all unified in a single model.
Dear fellow scholars, this is two minute papers with in this piece of work, we seek to control digital characters in real time. It happens the following. We specify a target trajectory and the algorithm has to synthesize a series of motions that follows that path to make these motions as realistic as possible. This is typically accomplished by unleashing a learning algorithm on a large database that that contains a ton of motion information. Previous techniques did not have a good understanding of these databases and they often synthesized motions from pieces that corresponded to different kinds of movements. This lack of understanding results in stiff, unnatural output motion. Intuitively, it is a bit like putting together a sentence from a set of letters that were cut out one by one from different newspaper articles. It is a fully formed sentence, but it lacks the smoothness and the flow of a properly aligned piece of text. This is a neural network based technique that introduces a phase function to the learning process. This phase function augments the learning with the timing information of a given motion. With this phase function, the neural network recognizes that we are not only learning periodic motions, but it knows when these motions start and when they end. The final technique takes very little memory, runs in real time and it accomplishes smooth walking, running, jumping and climbing motions and so much more over a variety of terrains with flying colors. In a previous episode we have discussed a different technique that accomplished something similar with a low and high level controller. One of the major selling points of this technique is that this one offers a unified solution for terrain traversal with using only one neural network. This has the potential to make it really big on computer games and real time animation. It is absolutely amazing to witness this and be a part of the future. Make sure to have a look at the paper which also contains the details of a terrain fitting step to make this learning algorithm capable of taking into consideration a variety of obstacles. I would also like to thank Claudio Panacci for his amazing work in translating so many of these episodes to Italian. This makes two minute papers accessible for more people around the globe. And the more people we can reach, the happier I am. Thanks for watching and for your generous support and I'll see you next time.
