Augmented Virtuality Input Demonstration Refinement Improving Hybrid Manipulation Learning for Bin Picking

Abstract

Beyond conventional automated tasks, autonomous robot capabilities aside human cognitive skills are gaining importance in industrial applications. Although machine learning is a major enabler of autonomous robots, system adaptation remains challenging and time-consuming. The objective of this research work is to propose and evaluate an augmented virtuality-based input demonstration refinement method improving hybrid manipulation learning for industrial bin picking. To this end, deep reinforcement and imitation learning are combined to shorten required adaptation timespans to new components and changing scenarios. The method covers initial learning and dataset tuning during ramp-up as well as fault intervention and dataset refinement. For evaluation standard industrial components and systems serve within a real-world experimental bin picking setup utilizing an articulated robot. As part of the quantitative evaluation, the method is benchmarked against conventional learning methods. As a result, required annotation efforts for successful object grasping are reduced. Thereby, final grasping success rates are increased. Implementation samples are available on: https://github.com/FAU-FAPS/hybrid_manipulationlearning_unity3dros