Towards Real-Time Simulation of Soft Robots with Contacts using a Method of Hybrid Hyper-Reduction

Soft robotics has emerged as an important part of robotics in recent years. Soft robots have an inherent view of contacts that is dramatically different from traditional rigid robots. Indeed, for rigid robots, contacts are either forbidden to avoid damage to the robot, the environment and humans, or precisely controlled for locomotion or interaction with an object. For soft robots, contacts may happen without damage, and when interacting with an object, local deformations allows for smoother interactions and potentially better performance. These prospects make soft robots attractive for tasks such as grasping. Fast finite element simulation is very useful for control and design. However, simulating collision adds a major numerical cost as it requires first a collision detection algorithm to detect collisions, and most importantly, it requires solving a constrained problem to avoid inter-penetrations and compute contact forces. When the number of contact points is large, this computation slows down the simulation dramatically. In this paper, we apply a hybrid hyper-reduction method to alleviate the FEM cost, the collision detection as well as the contact response computation. The deformations are computed in a low-dimensional subspace computed from offline experiments. The mechanical matrices are reduced through a method of hyper-reduction and the collision model is reduced following a hybrid reduction strategy. We show good agreement between original and reduced simulation while speeding up dramatically the computation. We first apply the method in simulation on a soft bouncing ball to explain the method. We then show an example with a soft gripper. The method is generic and can be used for control, design or learning algorithms.