A URI graduate student’s paper was recently awarded “Best Student Paper Finalist” for the 2023 Advanced Intelligent Mechatronics conference. Emadodin Jandaghi and Xiaotian Chen’s work was overseen by Associate Professor and Advisor Chengzhi Yuan.
As the name suggests, soft robotics involves using flexible materials in an increasingly wide variety of uses, from medical applications to underwater work. However, controlling them is difficult because they can be of infinite dimension. That, according to Emadodin Jandaghi, is why finding a way to model them is important. Without accurate detection of faults, or unexpected obstacles, a given procedure may be interrupted.
“The paper talked about a new way to train a soft robot model using a kind of machine learning called deterministic learning,” Jandaghi said. When a robot is in use, it follows its programmed trajectory. If there is a fault detected, the given operation may come to a halt. That’s where deterministic learning would come in. “We tested some simulation problems that could negatively impact the success of the robot. We had to define a framework that would allow detection of specific faults and to differentiate between types of faults.”
Theoretically, a robot might help a doctor during surgery. It may encounter an organ that shouldn’t be there. The goal is to train the robot before the operation so that it knows to take some alternate action whenever it confronts such a problem. At the same time, soft robots are different from each other, making modelling a challenge: the very fact that they are soft may mean that they can’t follow a certain trajectory precisely.
“That’s why modeling them is very important,” Jandaghi said. “And after that, when we define a precise model, then we could work on default detection framework. That is what we proposed in this paper.”
“Ensuring safety and accurate movement is essential in the operation of soft robots. The ability to detect faults is a critical aspect. To achieve this, it’s necessary to comprehend the motion dynamics of these robots, which is challenging due to their soft materials and nonlinear behavior. This has motivated scientists to adopt machine-learning techniques. Our research team has developed a Radial Basis Function Neural Network that offers higher accuracy in dynamics modeling which is proved mathematically unlike other methods,” said Chengzhi Yuan, professor and advisor.
By Hugh Markey