Robotic untangling and disentangling of cables such as power cords, ropes, wires, and hoses has applications in decluttering and prevention of tripping hazards in the home and workplace. It can also serve as a prerequisite for downstream tasks such as knot-tying in robot assisted surgical suturing or wire harness assembly in electrical and automotive manufacturing. However, one-dimensional deformable objects are difficult to model analytically and challenging for robots to visually track and manipulate. This is due to their infinite- dimensional state space, tendency to self-occlude, complex dynamics, and visually homogeneous appearance. Meanwhile, recent advances in deep learning, computer vision, and design of action primitives provide powerful means to handle increasingly complex manipulation tasks. Leveraging these advances, we present novel algorithms for single and multiple cable untangling and disentangling with learning-based perception.

First, we learn to untangle knots that involve multiple overlapping segments in single cables from RGB image observations. We build on a high-level planner that parameterizes loosening actions via image keypoints and introduce a low-level controller for fine-grained manipulation. This controller (1) performs grasp sampling and refinement on cables and (2) monitors untangling progress to perform recovery interventions accordingly. Next, we extend this manipulation stack to disentangle multiple cables.

We evaluate the proposed systems on physical experiments using the da Vinci surgical robot. In single-cable experiments, the robot achieves successful untangling in 68.3% of single cable knots with dense, non-planar starting states across 60 trials, outperforming baselines by 50%. In the multi-cable setting, the robot disentangles cables with 80.5% success, while generalizing across unseen knots and both distinct and identically colored cables.