The Taichi Programming Language for Fast and Differentiable Simulations

Yuanming Hu
09.14.2021 12:30 to 13:30

IRB 4105

Also on Zoom- Efficient and differentiable simulations can help solve inverse problems, such as robot design and control. However, with existing tools, developing these simulators is not easy. In this talk, we will cover a series of works on the Taichi programming language, including its differentiable version. After introducing basic prior knowledge such as (forward) physical simulation and differentiable (backward) simulation, we will talk about the ChainQueen differentiable simulator (ICRA 2019), the Taichi programming language (SIGGRAPH Asia 2019), and the DiffTaichi differentiable programming language (ICLR 2020). Taichi and DiffTaichi can help significantly improve the performance, productivity, and portability of physical simulator development, allowing (differentiable) simulators to run on GPUs, mobile phones, and even websites. As an open-source programming language embedded in Python, Taichi is now developed by nearly 100 developers across the world and has been downloaded ~1 million times (