World Model Enhanced Embodied Intelligence for Deformable Object Manipulation of Dynamic Targets

Deformable object manipulation, due to its widespread applications in both daily life and engineering, has been a popular research topic in robotics field. However, current methods always suffer from high computational complexity and poor interpretability. In this paper, we propose an iterative perception-gradient strategy for dynamic-target deformable object manipulation. The proposed policy refers to the idea of world model and constructs a perception environment internally for planning and decision-making, which integrates a perception dynamics network based on Transformer and Neural ODE architectures, a perception gradient estimation module, and an iterative policy to determine the robot's optimal actions. Numerous experimental results and statistical analyses demonstrate that the proposed policy maintains superior efficacy and reliability under various conditions and dynamic targets.