Soft Robot Control With a Learned Differentiable Model

Soft robots are inherently safe and comply readily to their environment. They are therefore exciting for applications like search and rescue or medicine, which involve a high degree of uncertainty, and require interacting with humans. However, the best way to model and control soft robots largely remains an open question. One promising approach is to leverage physically-based modeling techniques such as the finite element method. However, such techniques are inherently limited by their physical assumptions. Indeed, real-world soft robots are often made from unpredictable materials, using imprecise techniques. Data-driven approaches provide an exciting alternative, as they can learn real-world fabrication defects and asymmetries. In this paper we present our first investigation into using machine learning to do soft robot control. We learn a differentiable model of a soft robot’s quasi-static physics, and then perform gradient-based optimization to find optimal open-loop control inputs. We find that our learned model captures phenomena that would be absent from an idealized physically-based simulation. We also present practical techniques for acquiring high-quality motion capture data, and observations the effect of network complexity on model accuracy.