Hierarchical control using networks trained with higher-level forward models.

Area of Science:

• Robotics and Control Systems
• Artificial Intelligence
• Machine Learning

Background:

• Complex tasks often exceed the capabilities of single-level controllers.
• Existing control systems may struggle with dynamic environments and intricate task requirements.

Purpose of the Study:

• To propose and develop a hierarchical network control approach for complex tasks.
• To enable a low-level controller to manage a system (plant) and a higher-level controller to manage the low-level controller for enhanced capabilities.

Main Methods:

• A hierarchical control architecture with multiple levels of controllers.
• Training neural networks by minimizing cost functions based on optimal associations.
• Utilizing forward models, including a model of the low-level controller, for predicting command consequences during training.

Main Results:

• Successfully directed an articulated truck through an environment with obstacles.
• Trained networks demonstrated rapid response capabilities post-training.
• The hierarchical system effectively broke down complex tasks into manageable subtasks.

Conclusions:

• The hierarchical approach offers a scalable solution for complex task control.
• The system's modularity allows for adaptation to new tasks or increased complexity without full retraining.
• This method provides a robust framework for advanced robotic control and automation.