Learning for Deformable Linear Object Insertion Leveraging Flexibility Estimation from Visual Cues

TL;DR

This paper proposes a two-stage manipulation approach consisting of a material property (e.g., flexibility) estimation and policy learning for DLO insertion with reinforcement learning and trains a policy conditioned on the estimated flexibility to perform challenging DLO insertion tasks.

Abstract

Manipulation of deformable Linear objects (DLOs), including iron wire, rubber, silk, and nylon rope, is ubiquitous in daily life. These objects exhibit diverse physical properties, such as Young$'$s modulus and bending stiffness.Such diversity poses challenges for developing generalized manipulation policies. However, previous research limited their scope to single-material DLOs and engaged in time-consuming data collection for the state estimation. In this paper, we propose a two-stage manipulation approach consisting of a material property (e.g., flexibility) estimation and policy learning for DLO insertion with reinforcement learning. Firstly, we design a flexibility estimation scheme that characterizes the properties of different types of DLOs. The ground truth flexibility data is collected in simulation to train our flexibility estimation module. During the manipulation, the robot interacts with the DLOs to estimate flexibility by analyzing their visual configurations. Secondly, we train a policy conditioned on the estimated flexibility to perform challenging DLO insertion tasks. Our pipeline trained with diverse insertion scenarios achieves an 85.6% success rate in simulation and 66.67% in real robot experiments. Please refer to our project page: https://lmeee.github.io/DLOInsert/

Figures (9)

• Figure 1: Flexibility-aware DLO insertion. The real DLO insertion demonstration (lower) and corresponding simulation motion (upper) with estimated flexibility of the real DLOs.
• Figure 2: Pipeline for flexibility-aware DLO insertion in real experiment.We start by asking the robot to grasp the testing DLO using a predefined pose and sample particles on the DLO with image prepossessing techniques. The DLO states are input to the GNN-based estimation module to predict flexibility. Taking flexibility, DLO states and ring configuration as input, the policy can output a predicted grasping point and a control trajectory for insertion.
• Figure 3: Predefined grasping poses for DLO flexibility estimation in simulation (left) and real experiment (middle). The right image shows the DLOs used for the real experiment.
• Figure 4: Distance reward visualization.$\mathbf{d}_{floor}$ contributes negatively to reward when the DLO is not inserted. After insertion, a positive distance reward is given when the DLO is halfway or completely through.
• Figure 5: Action space visualization. The upper right square shows the range of $\mathbf{p}_{p}^{0}$ and the lower left square shows the range of $\mathbf{p}_{p}^{T}$. The rotation axis $\hat{\mathbf{n}}$ is pointing outward.