Deformation Configuration Estimation for Soft Continuum Robot Utilizing Seq2Seq Learning

Abstract

Inspired by biological tentacles, soft continuum robots exhibit the potential for navigating through narrow spaces and operating in complex environments, offering extensive application possibilities. However, owing to their inherent compliance, soft continuum robots may undergo unpredictable deformations in complex environments, leading to alterations in the whole-body configurations and diminished control precision. To address the problem, this letter employs sequence-to-sequence (Seq2Seq) learning to estimate the deformation of a tendon-driven continuum robot under multi-point contact. We also introduce a streamlined approach utilizing self-organizing mapping (SOM) to obtain ground truth data for training purposes and design dynamic loss functions for two-stage training, thereby facilitating neural network optimization in terms of both speed and precision. Furthermore, a soft continuum robot with two actively controlled degrees of freedom made of silicone is fabricated to verify the performance of the proposed method. The results show a shape estimation error of 2.94 mm (1.23% of the robot length).