A hexapod walker using a heterarchical architecture for action selection.

Area of Science:

• Robotics
• Artificial Intelligence
• Biomimetic Systems

Background:

• Controlling complex robotic systems, like hexapods with 22 degrees of freedom (DoFs), presents significant challenges in dynamic environments.
• Simple locomotion tasks require sophisticated state selection mechanisms to manage variations in velocity and external disturbances such as uneven terrain or leg slippage.

Purpose of the Study:

• To develop and demonstrate a novel control system for a six-legged walking machine capable of managing multiple complex behaviors.
• To enable adaptive locomotion, including forward/backward walking and tight curve negotiation, in cluttered or unpredictable environments.

Main Methods:

• A recurrent neural network (RNN) composed of motivation units was employed to manage attractor states.
• Decentralized memory elements were controlled by the RNN, integrating environmental feedback for adaptive state selection.
• A heterarchical network architecture facilitated the selection of diverse behavioral state combinations.

Main Results:

• The proposed RNN-based system successfully orchestrated various locomotion behaviors, including different walking speeds, backward locomotion, and negotiation of tight curves.
• The modular and heterarchical architecture demonstrated effective neural reuse, allowing adaptation to changing internal and external conditions.
• The system exhibited emergent properties, showcasing a holistic control approach beyond individual module contributions.

Conclusions:

• The developed RNN control system provides a robust solution for adaptive locomotion in complex robotic systems.
• The modular and heterarchical design offers a flexible framework for neural reuse and adaptation.
• This approach lays the groundwork for future cognitive robotic systems capable of advanced planning and navigation.