Robot Learning of Assembly Tasks from Non-expert Demonstrations using Functional Object-Oriented Network

摘要

Robot Learning from Demonstration (RLfD) is a research field that focuses on how robots can learn new tasks by observing human performances. Existing RLfD approaches mainly enable robots to repeat the demonstrated tasks by mimicking human activities, which usually requires efficient demonstrations for human experts. This paper proposes a new Function Object-Oriented Network (FOON) based approach to make robots learn and optimize assembly tasks from non-expert demonstrations. It first proposes an assembly FOON construction approach with automatic subgraph creation and merging algorithms to extract information from multiple non-expert demonstrations. It then proposes an assembly task tree retrieving approach with a robot execution optimization process to make the robot learn and generate the best possible task execution plan from the constructed FOON. The proposed approaches are validated through experiments with a dual-arm YuMi robot and the experimental results illustrate the effectiveness and advantages of the proposed approach.