哈尔滨工业大学（深圳）学术讲座

演讲人Speaker：赵琳

题目Title: Meta-Learning for Robot Control via Bilevel Optimization

时间Date：2023年 6月 23日       Time：上午 10：00 ~ 11：00

地点Venue： L 栋 216  室  

内容摘要Abstract： 

Robot estimation, planning, and control are often solved via optimization problems. Their solutions depend on various hyper-parameters including weighting matrices, reference waypoints, horizon lengths, etc. Real-time adaptation of these hyper-parameters is crucial for autonomous operations in complex and dynamic environments. We develop efficient algorithms for meta-learning these hyper-parameters modeled by deep neural networks (DNNs), achieving fast online adaptation in challenging robot control tasks. They are essentially solving a bilevel optimization problem, with the inner optimization problem subject to dynamic constraints. We present two interesting quadrotor control tasks solved in this way. One is robust flight control, where we developed a neural moving horizon estimator for accurate disturbance/model uncertainties estimation. The other is SE(3) path planning, where DNNs are trained to generate a reference waypoint to facilitate agile maneuvers. Both tasks demonstrate effective online adaptation and superior control performance.

个人简介(About the speaker):

赵琳，博士，新加坡国立大学助理教授，博士生导师。2006~2012 年在哈尔滨工业大学控制科学与工程系获得本科和硕士学位。2012~2017年在美国俄亥俄州立大学获得电气及计算机工程博士学位和数学硕士学位。2018至2020在美国安波福匹兹堡科技中心任研发科学家研究自动驾驶技术。2020年4月加入新国立电气及计算机系任助理教授。研究兴趣为控制理论，强化学习，以及在机器人中的应用。 曾担任2022年第17届IEEE International Conference on Control and Automation 会议程序联席主席， 担任第62届IEEE Conference on Decision and Control宣传联席主席, 以及Springer Journal of Systems Science and Complexity青年编委会。