Master's Thesis Research

I am currently a graduate research assistant in the Robotics and Dynamics (RaD) Lab at BYU. Our research focuses on modeling and control of soft robotics.

Currently, soft robots tend to be inaccurate in control tasks, but are useful when handling delicate objects or interacting with the environment. In the future we hope to have soft robots that can be used in assisted living, agriculture, and a number of other tasks that will improve human life. My prospectus has a complete outline of what I hope to accomplish throughout my graduate degree.

Publications

MoLDy: Open-source Library for Data-based Modeling and Nonlinear Model Predictive Control of Soft Robots submitted to RoboSoft 2024 (Primary Author)

PneuDrive: An Embedded Pressure Control System and Modeling Toolkit for Large-Scale Soft Robots submitted to RoboSoft 2024 (Co-author)

MoLDy: Open-Source Library for Data-based Modeling and Nonlinear Model Predictive Control of Soft Robots

My work on this paper was focused on an open-source toolkit (called MoLDy or Modeling of Learned Dynamics) for modeling and control of soft robots. The models are neural network based, with options to run hyperparameter optimizations based on hardware or analytical data. Once a model is trained the toolkit has a nonlinear model predictive controller that can run on the GPU for fast, real-time control. I have successfully created learned models that perform well in control for multiple systems, in simulation and on hardware. The next steps with this research is to use the learned models to perform complex, dynamic tasks on more complex hardware, as is shown below.

I had the chance to present my work at the RoboSoft 2024 conference in San Diego.

Transfer Learning towards effective control

My current work is building upon the open source repository, MoLDy to create a learned dynamic model of a full arm on the robot shown to the right. This robot is designed to complete "hugging" and grasping tasks. With the extra joints on the arm, the complexity in the model also scales. My approach is to use transfer learning to enable sim to real learning, reducing the amount of data that needs to be collected. The resulting joint-level controller can then be used in higher-level task control, such as reaching, manipulation, etc.