Development of a Learning Capability in Virtual Operator Models

This research developed methods for a virtual operator model (VOM) to learn the optimal control inputs for operation of a virtual excavator. Virtual design, used to model, simulate, and test new features, has often been limited by the fidelity of the virtual model of human operators. Human operator learns, over time, the capability, limits, and control characteristics of new vehicles to develop the best strategy to maximize the efficiency of operation. However, VOMs are developed with fixed strategies and for specific vehicle models (VMs) and require time-consuming re-tuning of the VOM for each new vehicle design. Thus, there typically is no capability to optimize strategies, taking account of variation in vehicle capabilities and limitations. A VOM learning capability was developed to optimize control inputs for the swing-to-pile task of a trenching operation. Different control strategies consisted of varied combinations of speed control, position control, and coast. A genetic algorithm (GA) was used to search for the best strategies and transition parameters to operate two vehicle design iterations. In the first design iteration, a combination of speed control resulted in the smallest time-to-swing from the trench to the pile. For the second VM design, a combination of speed control and coast resulted in the shortest time. This capability to learn the best control strategies for a new vehicle assists vehicle developers by avoiding time-consuming tuning of a VOM for each new vehicle design. In addition, the resulting VOM better represents the expert behavior typical of an operator who has operated the machine long enough to learn the best control strategy. This increases the utility and practicality of model-based design using closed-loop simulation of operators and new vehicle designs.