Deep Hybrid Models: Infer and Plan in a Dynamic World.

Area of Science:

• Cognitive Science
• Robotics
• Neuroscience

Background:

• Complex task planning often involves dynamic and hierarchical relationships.
• Traditional optimal control relies on cost function optimization.
• A biologically inspired alternative frames planning and control as an inference process called active inference.

Purpose of the Study:

• To present an active inference approach for complex task planning.
• To exploit discrete and continuous processing for enhanced planning capabilities.
• To extend planning as inference research and offer an alternative to optimal control.

Main Methods:

• Developed a deep hybrid model integrating discrete and continuous processing.
• Represented potential body configurations relative to objects.
• Utilized hierarchical relationships for flexible body schema expansion and tool use.
• Defined potential trajectories for inferring and planning with dynamic elements.

Main Results:

• Evaluated the model on a habitual task: reaching a moving object after picking a moving tool.
• Demonstrated the model's ability to handle the task under various conditions.
• Showcased the model's capacity for flexible planning and adaptation.

Conclusions:

• The proposed active inference approach effectively tackles complex, dynamic tasks.
• The deep hybrid model provides a novel framework for planning as inference.
• This work advances an alternative direction to traditional optimal control in robotics and AI.