ICON Seminar in Robotics: Prof. Changliu Liu (CMU)

In this talk, I will discuss our recent work on safe control and learning for effective human-robot collaboration. I will first introduce safe control methods using energy-function-based methods, then discuss how to combine them with learning controllers where an explicit analytical dynamic model of the system is usually not available (especially in human-robot systems). These safe control methods will enable safe reinforcement learning with zero training time violation. Then I will discuss about methods to robustly learn models to predict human behaviors. The key challenge we need to address is the distribution shift between the offline collected human behavioral data and the online measured human behaviors. To mitigate the distribution shift, we introduce two methods: online model adaptation, and offline verification-guided data augmentation. These methods have been applied to facilitate human-robot collaboration in industrial assembly tasks. I will conclude the talk with future visions on how to effectively deploy human-robot systems on factory floors.