Simulation and Manipulation of Cables in Isaac Sim: A Framework for Bimanual Robotic Control with Reinforcement Learning

This thesis presents a framework for the simulation and manipulation of deformable linear objects (DLOs), focusing on cables, using Isaac Sim. The research investigates three distinct cable modeling approaches: the mass-spring-damper method, Isaac Sim’s built-in deformable object model, and the PyElastica framework. After a comparative analysis, the PyElastica model was selected for its superior stability in simulating stiff deformable objects, such as cables, under dynamic conditions. Reinforcement learning (RL) was applied to train agents for two tasks using bimanual manipulation with Franka Emika Panda robotic arms. The first task, a foundational reach task, involved precise control of the robots’ end-effectors, achieving rapid convergence and demonstrating efficient learning within a simple control environment. The second task, an advanced cable manipulation task, required coordinated bimanual manipulation of deformable cables, significantly increasing complexity due to the cable's dynamic behavior. Using 2000 parallel simulation environments, the proposed framework enabled scalable RL training. The reach task achieved optimal performance in 222 iterations over 25 minutes, while the cable manipulation task required 272 iterations and five days of training due to the computational intensity of cable simulations. The results validate the framework's effectiveness for deformable linear object manipulation, showcasing the potential of reinforcement learning for solving complex robotic tasks.