Soft robotics, robotic manipulation, and control theory
- Controller-based Reinforcement Learning in Robotics Manipulation @ The Towards Autonomous Robots and Systems (TAROS) Conference 2023 / CDT Annual Conference / Joint Robotics CDT Conference – September 2023
I was born in China and moved to Japan when I was 12. Before joining the AgriFoRwArdS CDT community, I finished a bachelor’s degree in Mechanical Engineering at Osaka University. My research experience started when I was an exchange student at UC Berkeley, I was involved in research of argon power cycle at Combustion Modeling Lab. After taking several classes about robotics and control, I got interested and found opportunities to conduct work relating to autonomous driving including performance analysis and prediction algorithm implementation at Berkeley. For my undergraduate thesis, I developed a multi-robot cooperative transportation system that uses flexible tactile sensors.
I chose the AgriFoRwArdS CDT not only because the challenging and essential topics in robotics are addressed but also because of the positive environmental impact it could create through the application to agriculture. I got excited about moving to Lincoln because of the amazing view and buildings in the city. I will be studying for my PhD at the University of Cambridge under the supervision of Dr. Fulvio Forni.
I enjoy traveling and hiking to explore both culture and nature in different places during my spare time. I would like to take advantage and visit various cities and mountains while in the UK. Fun fact, I climbed Mountain Fuji twice in one month in the summer of 2022.
Object manipulation using model-based reinforcement learning
Using model-based Reinforcement Learning (RL) methods, we tackle the complex and nonlinear control problem in robotic strawberry harvesting where strawberries are in a cluster. To address this issue, we decided to solve a similar problem from the control field: balancing a 2D inverted pendulum. Our approach involves developing a simulation of the problem in Mujoco and using deep RL combined with dynamics model to expedite the learning process through the acquisition of the discrepancy between the dynamics model and real-world interactions. As part of the problem, we validate a Xela sensor simulation by a dataset of pushing experiments with a real Xela tactile sensor.
Data-driven autonomous robotic food handling
The project will develop new reliable data-driven control algorithms for robotic manipulation. Technologies for low-cost handling of food products will be explored in this project. Some low-cost articulated robotic arms equipped with grasping end-effectors will be developed and tested with the reasonable speed, accuracy, and reliability.
The research will focus on reliable data-driven control algorithms for food manipulation. The student will develop adaptive impedance control through a path of increasing complexity, starting from basic energy-aware control algorithms to a reliable adaptive control framework. This will be paired with mechanical design and prototyping, with the goal of co-designing control algorithms and (compliant, tuneable) hardware. The research will be validated on a commercial robotic system provided by RT Corp.
The student will have full access to all teaching courses at undergraduate, master, and PhD level, in engineering and wider domains (communication, management, etc.). The student will also learn from a large body of activities at the Control Laboratory, at the Bio-Inspired Robotic Laboratory, and at the Observatory for Human-Machine Collaboration.
Yi’s PhD project is being carried out in collaboration with RT Corporation, under the primary supervision of Dr Fulvio Forni.