Differential mapping spiking neural network for sensor-based robot control

Area of Science:

• Robotics
• Computational Neuroscience
• Artificial Intelligence

Background:

• Robotic systems require sophisticated control mechanisms to interpret sensor data and execute motor commands accurately.
• Approximating sensorimotor maps is crucial for enabling robots to perform complex tasks, but traditional methods can be computationally intensive.
• Spiking neural networks (SNNs) offer a biologically plausible and potentially more efficient alternative for modeling neural processes.

Purpose of the Study:

• To propose and validate a spiking neural network (SNN) for approximating differential sensorimotor maps in robotic systems.
• To develop a control architecture that leverages SNNs for efficient and accurate robot navigation.
• To demonstrate the SNN's capability in handling noisy sensor data for real-time control.

Main Methods:

• A spiking neural network (SNN) architecture was designed with sensory input and motor output layers, utilizing Izhikevich neurons and spike timing-dependent plasticity.
• A motor babbling process was employed to encode sensor feedback (proprioceptive and exteroceptive) into the SNN.
• Network parameters were tuned using a proposed guideline and particle swarm optimization.
• The SNN model was applied to a vision-guided robot for a target reaching task.

Main Results:

• The SNN effectively approximated local Jacobian-like projections, relating sensor changes to motor changes.
• The proposed control architecture successfully guided the robot to reach targets, minimizing path deviations and execution time.
• The SNN demonstrated low data and neuron requirements for training due to its architecture and optimized parameters.
• The system proved capable of real-time control and handling noisy sensor readings.

Conclusions:

• Spiking neural networks provide an effective and efficient method for approximating sensorimotor maps in robotic systems.
• The developed SNN-based control methodology offers a biologically plausible approach for real-time robot navigation with reduced computational demands.
• The approach shows promise for robust robot control in the presence of sensor noise and limited training resources.