Differentiable Physics Simulation

Abstract

Differentiable physics simulation is a powerful family of new techniques that applies gradient-based methods to learning and control of physical systems. It enables optimization for control, and can also be integrated into neural network frameworks for performing complex tasks. We believe that differentiable physics simulation should be a key component for neural networks to bridge the gap between training performance and the generality to previously unseen real-world inputs. However, realizing a practical differentiable simulation is still challenging because of its high dimensionality and fragmented computation flow. In this paper, we motivate the importance of differentiable physics simulation, describe its current challenges, introduce state-of-the-art approaches, and discuss potential improvements and future directions.