Learning to generate articulated behavior through the bottom-up and the top-down interaction processes

Area of Science:

• Robotics
• Artificial Intelligence
• Computational Neuroscience

Background:

• Sensory-motor learning and behavior generation are crucial for autonomous systems.
• Existing models often localize behavioral primitives, limiting flexibility.
• Hierarchical approaches offer potential for more sophisticated control.

Purpose of the Study:

• To propose a novel hierarchical neural network architecture for sensory-motor learning and behavior generation.
• To investigate the self-organization and sequential combination of behavioral primitives.
• To demonstrate robust and flexible situated behavior generation in a real-world robotic system.

Main Methods:

• Developed a two-level hierarchical neural network with forward models operating at different time scales.
• Enabled parametric interactions between network levels (bottom-up and top-down).
• Tested the architecture on a real robot arm with a vision system for learning and generation tasks.

Main Results:

• Behavioral patterns were learned through self-organization of primitives at the lower level and sequential combination at the higher level.
• Primitives were represented in a distributed manner, contrasting with prior localized approaches.
• On-line planning demonstrated robust behavior generation amidst real-world noise and flexible adaptation to environmental changes.

Conclusions:

• The proposed hierarchical architecture facilitates effective sensory-motor learning and behavior generation.
• Distributed representation of primitives enhances flexibility compared to localized methods.
• The interplay between bottom-up (recall) and top-down (prediction) processes enables robust and adaptive situated behavior.