Efficient End-effector Co-Design by Demonstration for Deformable Fragile Object Manipulation

Kei IKEMURA, Yifei Dong, David Blanco-Mulero, Alberta Longhini, Li CHEN, Florian T. Pokorny

Keywords: Co-Design, Reinforcement Learning, Deformable Fragile Object Manipulation

Abstract:

Manipulating deformable and fragile objects remains a critical challenge in robotics due to their complex dynamics and susceptibility to damage. Existing approaches typically address either hardware design or control policies in isolation.
In this work, we present the first co-design framework that simultaneously optimizes both end-effector design and control for deformable and fragile object manipulation. Our key insight is incorporating human priors through demonstrations, guiding the search for high-performance designs and control policies while maintaining sample efficiency. Our approach integrates a compact design space using cage-based deformation, a differentiable inverse design process, and a reinforcement learning algorithm (RLPD) to efficiently explore the joint design-control space. We evaluate our approach in the challenging task of grasping silk tofu. Preliminary experiment results demonstrate that our co-designed end-effector significantly reduces damage compared to the original parallel-jaw gripper. This work highlights the potential of co-adaptive design and control for deformable fragile object manipulation tasks

Submission Number: 9