A new hybrid learning control system for robots based on spiking neural networks

Area of Science:

• Robotics and Control Systems
• Computational Neuroscience
• Machine Learning

Background:

• Traditional nonlinear controllers often require manual tuning and struggle with complex dynamics.
• Spiking Neural Networks (SNNs) offer bio-inspired, energy-efficient computation suitable for real-time control.
• Reinforcement learning provides a powerful framework for adaptive parameter optimization in control systems.

Purpose of the Study:

• To introduce a hybrid learning and control method integrating SNNs and reinforcement learning.
• To enable adaptive parameter tuning for nonlinear controllers like Fractional Order PID (FOPID) and Feedback Linearization.
• To evaluate the performance and efficiency of the proposed method in various robotic applications.

Main Methods:

• Nonlinear controllers were modeled as multi-input multi-output functions and replaced by SNNs.
• Dopamine-modulated spike-timing-dependent plasticity (STDP) was employed for reinforcement learning and synaptic weight adjustment.
• The hybrid SNN-based controller was implemented and tested on robotic systems including mobile robots, inverted pendulums, and manipulators.

Main Results:

• The SNN-based FOPID controller (SNNFOPID) achieved low errors (e.g., 0.01 m, 0.03 rad) on tested robotic systems.
• The proposed method demonstrated superior performance in terms of Integral Absolute Error (IAE) compared to other approaches.
• Significant reductions in training time were observed for SNN-based Feedback Linearization (SNNFL) compared to SNNFOPID on certain tasks.

Conclusions:

• The hybrid SNN reinforcement learning method offers an effective and accurate approach for tuning nonlinear controllers.
• The method shows high potential for hardware implementation due to its low energy consumption, speed, and accuracy.
• The proposed framework is generalizable to other controllers and robotic systems, paving the way for broader applications.