DexDLO: Learning Goal-Conditioned Dexterous Policy for Dynamic Manipulation of Deformable Linear Objects

Deformable linear object (DLO) manipulation is needed in many fields. Previous research on deformable linear object (DLO) manipulation has primarily involved parallel jaw gripper manipulation with fixed grasping positions. However, the potential for dexterous manipulation of DLOs using an anthropomorphic hand is under-explored. We present DexDLO, a model-free framework that learns dexterous dynamic ma- nipulation policies for deformable linear objects with a fixed- base dexterous hand in an end-to-end way. By abstracting several common DLO manipulation tasks into goal-conditioned tasks, DexDLO can perform tasks such as DLO grabbing, DLO pulling, DLO end-tip position controlling, etc. Using the Mujoco physics simulator, we demonstrate that our framework can efficiently and effectively learn five different DLO manipulation tasks with the same framework parameters. We further provide a thorough analysis of learned policies, reward functions, and reduced observations for a comprehensive understanding of the framework. I. INTRODUCATION Deformable linear object (DLO) manipulation, e.g. ropes, cables, and rods, is widely applicable in surgical theaters, offices, textile factories, and other industries [1], [2]. Current research in DLO manipulation largely relies on single or dual parallel pinch grippers or end-effectors attached to a fixed end-tip of the DLO [3], [4], [5], [6], [7], [8]. How- ever, without task-specific customization, such end-effectors cannot provide sufficient dexterity for DLO manipulation tasks like in-hand DLO sliding and DLO weight pulling (see Figure 2 (a), (b), and (c) respectively). On the other ’hand’, an anthropomorphic hand, as a versatile end-effector, has the potential to handle all the aforementioned tasks. There are three common practices in the above-mentioned traditional DLO manipulation methods: 1) quasi-static DLO manipulation with a near zero velocity, 2) fixed grasp on the DLO, and 3) customized end-effectors to execute a specific DLO task. Compared to a two-finger gripper or DLO end- tip fixed end-effector, an anthropomorphic hand can avoid relying on these requirements during DLO manipulation, as shown in Figure 1, where the anthropomorphic hand grabs a DLO to fetch its end-tip. The major technical challenges of dexterous DLO manipulation are: • Dynamic manipulation. Previous continuous control methods mostly manipulated a DLO in a quasi-static state, assuming the DLO’s velocity is near zero [3], 1Sun Zhaole and Robert B. Fisher are with the School of Informatics, University of Edinburgh, UK. Corresponding author: zhaole.sun@ed.ac.uk 2Jihong Zhu is with School of Physics, Engineering and Technology, University of York, UK Fig. 1: An example of goal-conditioned dexterous manipulation of a deformable linear object. The controlled point (the grey segment) on the DLO is manipulated to minimize the distance to its goal position (the hand palm) and to finally reach the goal position using a base-fixed dexterous hand. and this assumption excludes some scenarios with non- zero velocities and higher manipulation speeds. Though handling the complexity of dynamics during real-time manipulation of the DLO is very difficult, dexterous manipulation of DLOs can benefit, making it more adaptable to different tasks. • Changing grasping positions. Preventing the DLO from slipping or falling from the hand during chang- ing grasping positions is challenging for parallel jaw grippers. Dexterous hands offer a unique advantage for this task without having to place and regrasp the DLO, which is often unavailable without a supporting surface, e.g. when picking and placing a rope on a table [7]. • General end-effectors. Different end-effectors are of- ten used for different DLO manipulation tasks, includ- ing specialized end-effectors, e.g., a gripper with tactile sensors [9] for DLO following1 and a specially designed gripper for rolling flat cables [10]. Dexterous hands are general-purpose end-effectors suitable for various rigid object manipulation tasks. Chen et al. [11] have shown that reorienting long and thin rigid objects in hand is difficult. Thus, it is even more difficult to use dexterous hands to manipulate DLOs, which are highly deformable long and thin objects and have high DoFs. Some works have already explored one or two aspects, e.g., DLO following to change grasping positions within the gripper with tactile sensors [9] or specialized end-effectors [10] and DLO dynamic whipping with a fixed grasping position [4], [6], [12]. However, none have systematically studied the dexterous manipulation of DLOs to address all three challenges, as discussed in Section II. To address the mentioned challenges, we introduce DexDLO, a reinforcement learning-based framework for 1Here, ‘following’ is defined to mean sliding the DLO through the hand. 2024 IEEE International Conference on Robotics and Automation (ICRA 2024) May 13-17, 2024. Yokohama, Japan 979-8-3503-8457-4/24/$31.00 ©2024 IEEE 16009