Online Iterative Learning Enhanced Sim-to-Real Transfer for Efficient Manipulation of Deformable Objects

Deformable manipulation has attracted a lot of attention in the field of robotics, especially in medical applications. However, manipulating deformable objects faces various challenges, mainly including their complex dynamic properties and unpredictable nonlinear deformations. It is difficult to provide a basis for deformable object measurements without effective control methods that provide intelligent and accurate position control, and this research also provides a premise for deformable object measurements. To address these issues, this paper proposes an online iterative perception policy (IPP) method, which does not require large-scale deep network training. This method is able to perceive transformations through an iterative process, and achieve efficient and accurate control of deformable objects. Extensive experiments in the simulation environment and the real scene are conducted to validate the effectiveness and superiority of the proposed method, as well as to compare with advanced algorithms (linear-quadratic regulator (LQR), sliding mode control (SMC), model predictive control (MPC), and heuristic). The experimental results reveal that IPP outperforms other approaches in terms of convergence, stability, robustness and flexibility in both the simulation and real-world scenarios, regardless of textile properties or initial conditions.