Learning force-based manipulation of deformable objects from multiple demonstrations

Abstract

Manipulation of deformable objects often requires a robot to apply specific forces to bring the object into the desired configuration. For instance, tightening a knot requires pulling on the ends, flattening an article of clothing requires smoothing out wrinkles, and erasing a whiteboard requires applying downward pressure. We present a method for learning force-based manipulation skills from demonstrations. Our approach uses non-rigid registration to compute a warping function that transforms both the end-effector poses and forces in each demonstration into the current scene, based on the configuration of the object. Our method then uses the variation between the demonstrations to extract a single trajectory, along with time-varying feedback gains that determine how much to match poses or forces. This results in a learned variableimpedance control strategy that trades off force and position errors, providing for the right level of compliance that applies the necessary forces at each stage of the motion. We evaluate our approach by tying knots in rope, flattening towels, and erasing a whiteboard.

Model-based reinforcement learning approach for deformable linear object manipulation

“…al. [6] employed machine learning for registering demonstrations to a new situation. More precisely, at first, multiple demonstrations were shown to the robot.…”

“…(6). Note that cov(x x x t , y y y t ) and cov(y y y t , x x x t ) can be analytically computed for the given policy.…”

“…On the other hand, machine learning has been employed to achieved DLO manipulation by means of a shown demonstration in advance [6]. This machine learning-based ap- proach avoids the need for a high-dimensionality DLO model by focusing on learning by registering demonstrations into a new situation.…”

Reinforcement learning for improving imitated in-contact skills

“…In this paper, we study how to improve an imitated wood planing task which exhibits an extremely complex dynamic interaction; every cut will generate a new environment, which does not allow for generalization [8], [9], [10] or planning [11] of the planing task. Hence, we apply the trial and error framework of reinforcement learning (RL) for improving the performance of the imitated planing skill.…”

Learning motions from demonstrations and rewards with time-invariant dynamical systems based policies

“…In the field of LfD, it has been proposed to derive stiffness variations via kinematic variability in demonstrated data [36]. Recent works have taken force measurements into account for estimating stiffness using weighted least squares [37] and least squares with platform specific priors on the stiffness parameters [38]. Another approach developed dedicated Human-Robot interfaces for the purpose of enabling stiffness variations to be easily taught to a robot [17].…”