A Differentiable Physical Simulation Framework for Soft Robots on Multiple-Task Learning

Yu Fang, Jiancheng Liu, Mingrui Zhang, Jiasheng Zhang, Yidong Ma, Minchen Li, Yuanming Hu, Chenfanfu Jiang, Tiantian Liu

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Differentiable Physics, Multiple-task Learning, Soft Robot Learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Learning multiple tasks is challenging for soft robots. Differentiable physics enables efficient gradient-based optimizations of neural network (NN) controllers for soft robot learning. However, existing work typically delivers NN controllers with limited capability and generalizability. We present a practical learning framework that outputs unified NN controllers capable of multiple tasks with significantly improved complexity and diversity. Our framework consists of a high-performance differentiable deformable bodies simulator supporting the material point method (MPM) and mass-spring systems, an automatic differentiation module that enables gradient-based optimizations, and a practical training module for soft robots on learning multiple locomotion tasks with a single NN controller. Using a unified NN controller trained in our framework, we demonstrate that users can interactively control soft robot locomotion and switch among multiple goals with specified velocity, height, and direction instructions. We evaluate our framework with multiple robot designs and challenging locomotion tasks. Experiments show that our learning framework, based on differentiable physics, delivers better results and converges much faster, compared with reinforcement learning frameworks. In addition, we successfully employed our framework on learning manipulation tasks, indicating the potential to extend our framework to tasks beyond locomotion.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8856