A limit-cycle self-organizing map architecture for stable arm control.

Area of Science:

• Computational Neuroscience
• Robotics
• Artificial Intelligence

Background:

• Conventional self-organizing maps (SOMs) use single-node, single-pattern representations, limiting information efficiency and robustness.
• Cerebral cortex activity exhibits oscillatory patterns, inspiring new neural network designs.

Purpose of the Study:

• To investigate limit cycle SOMs for controlling a robotic arm in inverse kinematics tasks.
• To develop a multi-map neural architecture integrating open-loop and closed-loop controls.
• To assess the generalization, accuracy, speed, smoothness, and robustness of this novel architecture.

Main Methods:

• Developed a multi-map neural architecture using limit cycle SOMs.
• Integrated open-loop and closed-loop control mechanisms.
• Simulated computer models for reach-and-hold tasks, including perturbations and map damage.
• Evaluated the architecture on a physical robot arm without further training.

Main Results:

• The limit cycle SOM architecture achieved accurate, fast, and smooth robotic arm movements.
• Demonstrated robustness against arm perturbations, map damage, and timing variations.
• Successfully controlled a physical robot arm, validating the simulation results.
• Showcased effective generalization capabilities for inverse kinematics.

Conclusions:

• Limit cycle SOMs provide an effective foundation for neural controllers.
• Harnessing non-fixed-point neural activity in SOMs enhances robotic control.
• This architecture offers a promising approach for robust and efficient robotic arm control.