Simulation of complex movements using artificial neural networks

Area of Science:

• Robotics and Artificial Intelligence
• Biomimetic Systems
• Control Theory

Background:

• Insect locomotion presents complex control challenges due to multi-joint coordination and environmental variability.
• Existing control systems often struggle with dynamic changes in terrain, system geometry, and multi-leg coordination.
• Developing robust and adaptive control for legged robots is crucial for versatile locomotion.

Purpose of the Study:

• To propose and evaluate a simulated network for controlling a six-legged, insect-like walking system.
• To investigate the efficacy of environmentally-coupled recurrent connections for locomotion control.
• To demonstrate a decentralized control approach that addresses inherent system non-linearities and external perturbations.

Main Methods:

• Development of a simulated six-legged walking system with internal recurrent connections.
• Implementation of a decentralized controller utilizing negative and positive feedback at the joint level.
• Analysis of the controller's performance under varying conditions, including curve walking, compliant substrates, and changing system geometry.

Main Results:

• A simple subnet effectively controls leg swing, utilizing environmental loops.
• The decentralized controller successfully manages the complex stance phase, coordinating 9-18 joints.
• The system demonstrates robustness against variations in leg stance combinations, speeds, substrate compliance, and geometry.

Conclusions:

• A decentralized, joint-level feedback controller can effectively manage complex insect-like locomotion.
• Exploiting the physical properties of the system is key to achieving robust and adaptive walking.
• This approach offers a potentially simpler and more effective alternative to traditional centralized control methods for legged robots.