Dynamic Bimanual Cloth Manipulation Via Dynamic Movement Primitives and Reinforcement Learning

Learning effective control strategies for de- formable object manipulation remains a major challenge in robotics, especially when tasks require fast, coordinated motions between multiple manipulators. In this work we address the problem of dynamic bimanual cloth manipulation, where two robotic arms must coordinate fast, fluid motions to place a cloth onto a surface. Our method uses Proximal Policy Optimization (PPO) with Dynamic Movement Primitives (DMPs) as the policy output, enabling smooth and parameterized trajectory generation. We compute rewards such as cloth height, corner alignment, and movement direction only at the end of each training episode, while the robot control continues at a higher frequency, and introduce a probing phase to obtain knowledge about cloth dynamics. We implement our approach in NVIDIA Isaac Sim with realistic cloth dynamics. Experiments show that this setup allows the robots to learn fast, coordinated bimanual cloth placement using only occasional reward feedback.