Lenna Kanehara
MIT EECS Undergraduate Research and Innovation Scholar
Autodiff Tools for Differentiable Physics Simulation
2021–2022
Electrical Engineering and Computer Science
- Graphics and Vision
Wojciech Matusik
Physics simulation has wide applications in computer graphics, mechanical engineering, and robotics. Particularly, differentiable physics simulation has become increasingly popular as the calculation of gradients allows for richer analysis of a physics system. However, deriving gradients in a simulator is labor-intensive and error-prone. Existing autodiff tools such as Pytorch are not designed for data structures used in physics simulation. My goal is to develop an autodiff tool specialized for differentiable simulation. Specifically, I will focus on automatic calculation of higher-order derivatives and differentiating implicit functions. With the tools I develop, I hope to unlock downstream applications in physics simulation, controller design, and inverse problems.
I am participating in SuperUROP to gain research experience in the field of computer graphics. I look forward to applying what I learned in 6.837 and 6.839 and contributing to the research in differentiable physics simulation. The project is exciting as it intersects my interest in computer graphics, programming, and physics. I hope to deepen my understanding in all these fields as well as improve my research and communication skills.