Soft Robot Workspace Estimation via Finite Element Analysis and Machine Learning

Soft robots with compliant bodies offer safe human–robot interaction as well as adaptability to unstructured dynamic environments. However, the nonlinear dynamics of a soft robot with infinite motion freedom pose various challenges to operation and control engineering. This research explores the motion of a pneumatic soft robot under diverse loading conditions by conducting finite element analysis (FEA) and using machine learning. The pneumatic soft robot consists of two parallel hyper-elastic tubular chambers that convert pneumatic pressure inputs into soft robot motion to mimic an elephant trunk and its motion. The body of each pneumatic chamber consists of a series of bellows to effectively facilitate the expansion, contraction, and bending of the body. The first chamber spans the entire length of the soft robot’s body, and the second chamber spans half of it. This unique asymmetric design enables the soft robot to bend and curl in various ways. Machine learning is used to establish a forward kinematic relationship between the pressure inputs and the motion responses of the soft robot using data from FEA. Accordingly, this research employs an artificial neural network that is trained on FEA data to estimate the reachable workspace of the soft robot for given pressure inputs. The trained neural network demonstrates promising estimation accuracy with an R-squared value of 0.99 and a root mean square error of 0.783. The workspaces of asymmetric double-chamber and single-chamber soft robots were compared, revealing that the double-chamber robot offers approximately 185 times more reachable workspace than the single-chamber soft robot.