Gaussian Process Dynamics Models for Soft Robots with Shape Memory Actuators

Abstract

Efficient and tractable dynamics models of soft robots are needed for many purposes. For soft robots powered by shape memory actuators, models must simultaneously capture actuator dynamics alongside material deformations. This article examines one possible modeling framework using Gaussian Process (GP) regression, applied in two settings. First, we attempt predictions of bending deformation in a low-dimensional hardware task of a single limb with one shape memory alloy (SMA) actuator, where the SMA temperature is estimated in open loop. Second, we consider locomotion in a simulation of a rolling robot, where the actuator state is known perfectly but the state space is high-dimensional and hybrid. Our locomotion tests examine the applicability of standard model selection choices for a GP, particularly the commonly-used squared exponential (SE) kernel, by studying the hyperparameter optimization problem. We cross-validate one-step-ahead dynamics predictions for both robots. All results have significant noise, but for the single limb in bending, the GP predictions are sufficient to simulate motions over a time horizon of 30 seconds. With further work to address issues with data quality, underfitting, and other sources of error, these dynamics models have the potential to assist in trajectory generation and control for soft robots.