Coupled Multiple Dynamic Movement Primitives Generalization for Deformable Object Manipulation

Abstract

Dynamic Movement Primitives (DMP) are widely applied in movement representation due to their ability to encode tasks using generalization properties. However, the coupled multiple DMP generalization cannot be directly solved based on the original DMP formula. Prior works provide satisfactory performance for the coupled DMP generalization in rigid object manipulation, but their extension to deformable objects may degrade due to the intrinsic uncertainty of the deformable model structure and parameters. This paper introduces an adaptive term to replace the fixed term to couple multiple DMP generalizations and model the deformable object using the classic mass-spring-damper model. Based on the modeling, the manipulation of a deformable object can be treated as a second-order system, which provides additional implementation flexibility and robustness in deformable object transportation. To validate the proposed method, we perform extensive simulations for cooperatively transporting a rope and a deformable thin film, imitating the manipulation with a semi-ellipse trajectory and M-shape trajectory. We further implement our method on a dual-arm robot platform for rope manipulation with depth visual feedback. Both simulation and experiment results demonstrate satisfactory DMP generalization, collision avoidance, and configuration preservation.

A Practical Roadmap to Learning from Demonstration for Robotic Manipulators in Manufacturing

“…Teaching the robot about variations in object properties is essential for tasks where the characteristics of objects significantly impact the manipulation process. For instance, in material handling tasks, the robot needs to learn how to handle objects of different shapes, sizes, weights, and materials [85,88]. Demonstrations can be designed to showcase the manipulation of diverse objects, allowing the robot to generalize its learning across a range of scenarios.…”

“…In another context, [49] considers the assembly use cases and explores the idea that skillful assembly is best represented as dynamic sequences of manipulation primitives, and that such sequences can be automatically discovered by Reinforcement Learning. The authors in [88] extended DMPs to manipulating deformable objects such as ropes or thin films, to account for the uncertainty and variability from the model parameters of the deformable object. Another important aspect of learning is to learn factor which are in non-euclidean spaces.…”

Dynamic movement primitives in robotics: A tutorial survey

“…Prakash et al (2020) extended the real-time adaptation approach incorporating a fuzzy fractional-order sliding mode controller in order to efficiently and stably adapt the demonstrated DMP trajectory to fast movements, such as a ping pong swing. Recently, Cui et al (2022) presented a method for coupling multiple DMPs for modeling robot tasks for transportation tasks of deformable objects.…”

Multirobot Transport of Deformable Objects With Collision Avoidance

“…Those methods do not tackle the problem of explicitly controlling the deformation of the object in 3D, and they do not consider collision avoidance in their formulation. We can also find approaches based on Dynamic Movement Primitives (DMP) [10] or human-robot collaborative systems for transport of highly deformable planar materials [11], which do not test collision avoidance against dynamic obstacles.…”