Behavioral Learning in a Cognitive Neuromorphic Robot: An Integrative Approach.

Area of Science:

• Cognitive Robotics
• Neuromorphic Engineering
• Computational Neuroscience

Background:

• Integrating spiking neural networks (SNNs) with robots presents challenges but offers insights into cognitive architectures.
• Advances in neuromorphic processing and cognitive robotics enable complex SNN-robot integrations.
• Dedicated neural hardware is crucial for exploring SNNs in robotic learning.

Purpose of the Study:

• To develop and evaluate a learning system for object-specific attention using the iCub humanoid robot and SpiNNaker neuromorphic chip.
• To investigate how SNNs and learning rules can yield insights into neural architecture and learned behavior in a cognitive robotics context.
• To demonstrate a scalable, modular approach for building and testing SNNs for robotic applications.

Main Methods:

• Developed a scalable, structured, and modular spiking neural network architecture.
• Implemented a classical spike-timing-dependent plasticity (STDP) learning rule on selected connections.
• Introduced structural enhancements to the network to direct performance toward behaviorally relevant goals.
• Utilized the iCub humanoid robot and SpiNNaker neuromorphic chip for real-world task execution.

Main Results:

• The system demonstrated significant improvement in object-specific attention task performance through STDP.
• Behaviorally relevant STDP strongly contributed to positive learning (e.g., "do this") but less to negative learning (e.g., "don't do that").
• Structural enhancements to the SNN had a cumulative positive effect on performance.
• The combination of effects, rather than isolated properties, was key to achieving compelling, task-relevant behavior.

Conclusions:

• Spiking neural networks, when integrated with robotic platforms and appropriate learning rules, can achieve sophisticated, task-relevant behaviors.
• Neuromorphic hardware and cognitive robotics approaches facilitate the study of SNNs for understanding computation and behavior.
• The modular and scalable design allows for adaptation to new tasks and further investigation of learning mechanisms.