Dual-arm robot planning and control with DMPs for soft objects manipulation

Abstract

The trajectory design of manipulators for manipulating soft objects is still an unsolved problem. Different from traditional rigid objects, soft objects are fragile with different deformation behavior and dynamic physical properties. Therefore, grasping thin objects such as bed sheets or biofilms, are more preferable to be handled by multiple manipulators, rather than a single manipulator, for delicate operation. Considering the manipulator's self-collision and soft object physical characteristics, it is difficult to obtain a suitable trajectory that can coordinate multiple manipulators and be applied to a different similar task. Therefore, to achieve fast and accurate soft object manipulation, we have studied the trajectory design of multiple robotic arms (including the trajectory of manipulators and humans) for manipulating soft objects. In this thesis, we model the soft manipulation trajectory with the dynamic movement primitive (DMP) and couple the multiple DMPs as an integrated system based on the Spring-Mass-Damper model and formation control techniques to complete soft object manipulation tasks. Based on the Spring-Mass-Damper model, the aims of the multiple DMPs generalization and pattern preserving of the manipulated object can be completed at the same time. With the introduction of the adaptive coupling term, the coupled DMP system can be adaptively applied to different kinds of manipulation tasks. Different from the above Spring-Mass-Damper model, we proposed another method to couple multiple DMPs based on the formation control technique in which the whole DMP system is more flexible and effective, especially in applications where the robot arm number is larger than 4. To validate our methods, extensive simulation and real experiments based on a Bullet physics simulator, Kinova, and UR5 robot are tested and compared with the benchmark methods. As a natural extension of multi-robot soft manipulation, the cooperation between the robot and human in soft object manipulation is also researched. While the human agent serves as a leader, the trajectory planning, and optimization of whole manipulation (transportation) can be greatly simplified. However, these merits come along with extra-human disturbances such as human handshaking. This problem is especially severe in some precision applications. Hence, we proposed that the soft object model is firstly estimated by Least Square with an exponential forgetting algorithm, and the human hand disturbance is considered as extra noise for the model estimation and model update. The local model estimation is obtained in an online adaptive manner. Because of that, the soft model would be estimated more accurately so that the human and robot can corporately transport the soft object.