Sim-to-Real for Differentiable Projective Dynamics on Soft Robots: Meshing, Damping, and Actuation

Mathieu Dubied, Mike Yan Michelis, Andrew Everett Spielberg, Robert K. Katzschmann

Abstract:

An accurate, physically-based model of soft robots can unlock downstream applications in optimal control. But modeling soft robots has mostly involved coarse approximations that do not suitably describe the dynamics and deformations. The Finite Element Method (FEM) is a more expressive approach for modeling highly deformable structures such as dynamic, elastomeric soft robots. Recently, Differentiable Projective Dynamics (DiffPD) has been proposed as a fast and differentiable FEM; however, DiffPD lacks rigorous benchmarking against reality. In this paper, we compare virtual robot models simulated using DiffPD with measurements from their physical counterparts. In particular, we examine several soft structures with different morphologies: a clamped beam under external force, a pneumatically actuated soft robotic arm, and a soft robotic fish tail. We also advance DiffPD by proposing a predictive model for pneumatic soft robot actuation.
Through our case-studies, we provide strategies and algorithms for matching real-world physics in simulation, thus making DiffPD useful for robots in reality.