MIT researchers have developed a new tool that provides better control to animators in shaping their characters. The new technique works by generating mathematical functions, known as barycentric coordinates, that describe how 2D and 3D shapes in animations can move, stretch, and deform in space. By using these functions, an animator can tailor the movement of a character to fit their vision.
Existing methods often offer only one option for the barycentric coordinate functions for each character, forcing the animator to start from scratch if they want a different look. The MIT researchers’ approach allows more flexibility, letting artists select among various functions that correspond with different versions of animation “smoothness.”
The team’s approach also incorporated neural networks into the design process. This network architecture can output barycentric coordinate functions that meet all necessary constraints without requiring the animator to handle complex mathematical tasks. The researchers built constraints directly into the network, allowing the tool to generate valid solutions and simplifying the design process for the artist.
To overcome the complexity of modern cages (used to define character shapes), the researchers covered shapes with overlapping ‘virtual’ triangles, each forming a valid barycentric coordinate function. The neural network then determines how to combine these virtual triangles to create a more complex, yet smooth function.
The method allows for iterative experimentation. Artists can try one function, evaluate the resulting animation and modify the coordinates to get the desired effect. Demonstrations show how the tool could generate more naturally flowing animations compared to other methods, like making a cat’s tail curve smoothly instead of folding rigidly.
In addition to assisting in animation, the technique holds potential for use in diverse contexts including medical imaging, architecture, virtual reality, and computer vision systems for robots. The researchers aim to further refine the technique for real-time use and to accelerate the neural network for faster operation. The research was backed by the U.S. Army Research Office, the U.S. Air Force Office of Scientific Research, MIT-IBM Watson AI Lab, and other entities.
