Compensating for Unmodeled Forces using Neural Networks in Soft Manipulator Planning

Abstract

Soft manipulators made of deformable materials have great promise in applications that require additional flexibility and compliance; however, these characteristics also make them difficult to simulate accurately and quickly. The lack of a fast and accurate simulator prevents motion planners from generating feasible plans, which would enable soft robots to achieve more complex tasks, such as manipulation. In this work, we propose combining a simplified quasistatic model with a neural network that learns to compensate for unmodeled forces, such as friction and loads, in order to create a fast forward model for soft manipulators with multiple segments. We show that the resulting neural network model reduces average end effector position error by 62% compared to the quasistatic model, while still being fast enough for motion planning. We also incorporate this model into an RRT*-based planner and demonstrate that the plans generated using our model are more likely to be feasible when executed on hardware than plans generated with a simulator using the quasistatic model.