Task Learning, Intent Prediction, and Adaptive Blended Shared Control With Application to Excavators

Abstract

Human operators of construction equipment usually exhibit strong situational awareness, which enables them to execute tasks while handling unexpected uncertainties and adapting to environmental changes. Automatic control, on the other hand, is generally capable of executing repetitive and well-defined tasks more efficiently and precisely. In dynamic and uncertain environments, significant safety and performance benefits can be derived by collaboratively achieving the tasks via the blending of control inputs from the human operator and automatic control. The focus of this article is on developing efficient methods for such blending and its application to hydraulic excavators, which involves the developments in task learning, intent prediction, and human-machine shared control. We propose a new task learning method by segmenting the tasks with the operator primitive-based segmentation (OPbS) and clustering of subgoals via Bayesian nonparametric clustering with temporal ordering (BNPC/TO). We introduce a method for dynamically predicting the operator's intent in seeking a particular subgoal by proposing an empirical stochastic transition matrix (ESTM) and a dynamic angle difference exponential (DADE). We then propose a method for blended shared control with conflict awareness extended from dynamic angle difference. Finally, we apply our algorithms and evaluate the results on a scaled hydraulic excavator test platform for a typical earth-moving task with novice learning operators and a skilled demonstration operator.