Grasping 3D Deformable Objects via Reinforcement Learning: A Benchmark and Evaluation

Abstract :

Robotic manipulation of deformable objects is a challenging task that has been tackled with a variety of approaches. However, due to the highly difficult task of modeling the dynamics of deformable objects in a fast and accurate way, many real-world use cases remain unsolved. Recent advances in data-driven approaches like reinforcement learning (RL) promise that these methods push forward the envelope of feasibility in the field of deformable object manipulation. Despite the growing interest in this field, data-driven approaches mainly focus on the manipulation of 1D and 2D deformable objects like ropes and cloth. In this work, we present the benchmark DeformableGym to facilitate the evaluation of RL methods for grasping 3D deformable objects. We use a set of simulated benchmark environments to evaluate existing model-free state-of-the-art algorithms and investigate the main challenges and potential pitfalls of applying them in this challenging setting.

Files:

mdo-camera-ready.pdf

Links:

https://deformable-workshop.github.io/icra2023/
https://github.com/dfki-ric/deformable_gym