Iterative Residual Policy for Goal-Conditioned Dynamic Manipulation of Deformable Objects

• DOI: 10.15607/rss.2022.xviii.016
• Publication Date: 2022-07-02T22:49:15Z
• Authors: Cheng Chi; Benjamin Burchfiel; Eric Cousineau; Siyuan Feng; Shuran Song
• Keywords: FOS: Computer and information sciences; Computer Science - Robotics; 0209 industrial biotechnology; 02 engineering and technology; Robotics (cs.RO)

This paper tackles the task of goal-conditioned dynamic manipulation deformable objects.This is highly challenging due to its complex dynamics (introduced by object deformation and high-speed action) strict requirements (defined a precise goal specification).To address these challenges, we present Iterative Residual Policy (IRP), general learning framework applicable repeatable tasks with dynamics.IRP learns an implicit policy via delta dynamicsinstead modeling entire dynamical system inferring actions from that model, IRP predict effects action on previously-observed trajectory.When combined adaptive sampling, can quickly optimize online reach specified goal.We demonstrate effectiveness two tasks: whipping rope hit target point swinging cloth pose.Despite being trained only in simulation fixed robot setup, able efficiently generalize noisy real-world dynamics, new objects unseen physical properties, even different hardware embodiments, demonstrating excellent generalization capability relative alternative approaches.