Deformable Object Manipulation with a Tactile Reactive Gripper

Abstract

Deformable objects like cloth and cables are challenging for robots to manipulate due to their high-dimensionality and unpredictable dynamics. In previous work, Yu et. al (2019) [45] used a tactile sensor to estimate the pose of a cable within the grip while sliding along it. The authors used linear regression to model the cable sliding dynamics and used a linear quadratic regulator (LQR) controller to keep the cable centered within the grip. However, the underlying dynamics are not linear, so in this work, we explore controllers that take advantage of a non-linear underlying dynamics model. We use Gaussian process (GP) regression for the non-linear model which is used in three controllers in hardware experiments: (1) LQR with the GP model linearized about the target position and (2) time-varying LQR with the GP model linearized about the current state and (3) model predictive control with the full dynamics model and constraints on the state and input of our system over a finite horizon. We extend our framework for the more challenging task of cloth edge following by adjusting our hardware setup and developing a new perception system. We found that the time-varying LQR controller using the GP model performs similarly to the LQR controller with the linear regression model for following both cables and fabric edges.