Adaptive Control for Robotic Manipulation of Deformable Linear Objects with Offline and Online Learning of Unknown Models

Abstract

The deformable linear objects (DLOs) are common in both industrial and domestic applications, such as wires, cables, ropes. Because of its highly deformable nature, it is difficult for the robot to reproduce human's dexterous skills on DLOs. In this paper, the unknown deformation model is estimated in both the offline and online manners. The offline learning aims to provide a good approximation prior to the manipulation task, while the online learning aims to compensate the errors due to insufficient training (e.g. limited datasets) in the offline phase. The offline module works by constructing a series of supervised neural networks (NNs), then the online module receives the learning results directly and further updates them with the technique of adaptive NNs. A new adaptive controller is also proposed to allow the robot to perform manipulation tasks concurrently in the online phase. The stability of the closed-loop system and the convergence of task errors are rigorously proved with Lyapunov method. Simulation studies are presented to illustrate the performance of the proposed method.

Offline-Online Learning of Deformation Model for Cable Manipulation With Graph Neural Networks

“…Though the online model estimation methods are useful in certain tasks, it is still desirable to take advantage of the more powerful offline global model. As demonstrated in [2], the offline learning provides the initial guess of the linear Jacobian matrix, and in the online phase, the Jacobian matrix is further updated with an adaptive control law. Distinct from the sim-to-real methods introduced in the previous subsection, this combination does not require additional online data collection.…”

“…The sim-to-real gap can be instantaneously resolved as the robot executes. While the motivation of our proposed framework is similar to the motivation of [2], our algorithm is unique in several aspects: 1) we utilize GNN to capture the global model instead of providing the initial guess of the local Jacobian, 2) we online learn a residual model for refinement, 3) we construct an optimization-based MPC controller to obtain optimal robot motions within a long horizon.…”

“…Aside from the FEM models, data-driven approaches are often utilized to learn the non-linear deformation dynamics [2]- [4]. Recently, many works have attempted to use deep Fig.…”

Feel the Tension: Manipulation of Deformable Linear Objects in Environments with Fixtures using Force Information

“…For example, Berenson [23] introduced the concept of diminishing rigidity to define a Jacobian where points on the DLO closer to the grasped point are assumed to act more rigidly. Other approaches have been proposed where a Jacobian is learned using neural networks [24] or estimated through weighted least-squares [25]. In this work, we instead estimate a reduced graph model.…”

One-shot Learning Closed-loop Manipulation of Soft Slender Objects Based on a Planar Polynomial Curvature Model

“…Yet, the task is substantially more complex for deformable bodies due to their distributed nature. Machine learning algorithms have been widely used to approximate an endto-end representation of the object's behaviors [6], [7], [8] or directly a manipulation policy [9], [10], [11]. Similar approaches have been investigated in soft robotics [12], [13], [14].…”