VOR Adaptation on a Humanoid iCub Robot Using a Spiking Cerebellar Model

Area of Science:

• Robotics
• Computational Neuroscience
• Control Systems Engineering

Background:

• Spiking neural networks (SNNs) offer biologically plausible computation but present challenges in real-time (RT) control due to variable computation times.
• The cerebellum plays a crucial role in motor adaptation and learning, particularly in vestibulo-ocular reflex (VOR) tasks.
• Integrating complex neural models into RT robotic systems requires robust mechanisms to manage computational demands and ensure stability.

Purpose of the Study:

• To develop and evaluate a real-time control system for a robotic platform (iCub) using a spiking cerebellar model for vestibulo-ocular reflex (VOR) tasks.
• To address the challenge of nondeterministic computation time in SNNs by implementing an adaptive real-time supervisor module.
• To investigate how cerebellar neural computation and plasticity contribute to motor adaptation in VOR tasks.

Main Methods:

• Embedding a spiking cerebellar model within an adaptive real-time control loop operating the iCub robot.
• Implementing a real-time supervisor module to manage the SNN's computation time, employing countermeasures like halting or disabling features (e.g., STDP, spike propagation).
• Applying the neurorobotic setup to horizontal and vertical VOR adaptive tasks, widely used in neuroscience research.

Main Results:

• Demonstrated successful operation of the iCub robot in VOR tasks using the spiking cerebellar model.
• The RT supervisor module effectively managed computational overloads and ensured stable robot operation by adjusting simulation speed.
• The experimental setup allowed for the investigation of cerebellar neural activity and plasticity in mediating motor adaptation.

Conclusions:

• A two-stage learning process is necessary for effective VOR acquisition.
• The integration of a spiking cerebellar model with an RT supervisor module is a viable approach for neurorobotic control.
• Biological sensory-motor delays can be leveraged to buffer computational demands in real-time neurorobotic systems.