Long cables are widely found in domestic and industrial settings. These linear deformable objects form knots that hinder functionality and pose safety risks. Untangling them is crucial for restoring proper cable function and enabling downstream tasks like surgical suturing and automotive wire harness assembly. However, untangling long cables is challenging due to their infinite-dimensional state-space, self-occlusion, and knot formation. Longer cables are even more prone to knotting, and as one segment is untangled, the remaining segments, known as "slack," can create new knots.

To address these challenges, our research focuses on developing RGB perception and motion primitives, along with specialized gripper jaws, for efficient untangling of long cables. Our algorithm utilizes these advancements to iteratively untangle cables, achieving success rates of 67% for isolated knots and 50% for complex configurations.

In the project’s initial phase, we identified that failure primarily stems from drastic actions in uncertain states, resulting in irrecoverable cable states. We introduce novel metrics and cable-interacting actions that reduce perception uncertainty before untangling maneuvers. This integrated system accommodates variations in cable states, achieving an 83% success rate in untangling cables with one or two knots and a 70% termination detection success across these configurations. Our method demonstrates a 43% improvement in untangling accuracy and completes the task three times faster compared to the previous system.

Despite these improvements, the system still fails to untangle cables in out of distribution states. We develop a generalized perception approach using a novel cable state estimator, including a learning-based iterative tracer, crossing classifier, and crossing correction algorithm for semi-planar knots. Implementing this system leads to a 13% improvement in untangling success for complex knots. Additionally, our method demonstrates versatility by achieving an 81% success rate in multi-cable tracing. Furthermore, it handles variations in cable appearances, tracing cables not in the training dataset with an 85% success.




Download Full History