Located at the base of the brain, the cerebellum is responsible for the coordination, control and timing of movements. This structure allows us to walk run and perform other motor-related activities, such as throwing a baseball, without consciously having to orchestrate the individual movements. These cerebellum-driven motor activities are some of the most difficult to reproduce in the robot population.

Japanese researchers Tadashi Yamazaki of the University of Electro-Communications in Tokyo and Jun Igarashi at Okinawa Institute of Science and Technology Graduate University in Okinawa used NVIDIA GPUs to create a 100,000 neuron simulation of the human cerebellum. This is one of the largest simulations of its kind. They put their model to the test by linking it to a robot that relies on the virtual cerebellum to hit a ball.

This field of study, known as “biomimetic” robotics, relies on biological systems to inspire the design and engineering of materials and machines with the aim of developing a more compliant and robust class of robots than today’s current generation.

The research duo has written a paper describing their work and outlining their goals. Originally, they designed the large-scale network model to study the underlying mechanisms of cerebellum motor control, but it soon became apparent that the virtual cerebellum could be used to help robots interact with and respond to their environments in a more natural way, potentially advancing one of the most vexing problems in robot research.

One of the biggest challenges to real-time neural modeling is simulation speed. A CPU-only solution took 98 seconds to generate a response to a one-second stimulus. Owing to its massive parallel computing capability, the GPU-based solution, was one-hundred times faster, making it a suitable candidate for use in real-world scenarios. The parallel implementation of the “Realtime Cerebellum (RC)” platform was carried out using CUDA, NVIDIA’s unified software development environment for GPU programming.

To test their work, the researchers connected RC to a humanoid robot that they had built. Using RC as a real-time adaptive controller, the robot learns to hit a small plastic ball with a round racket. (See video below.)

What makes this work all the more significant is that it was done with inexpensive off-the-shelf hardware. The model used a PC equipped with a single GeForce GTX 580 NVIDIA GPU.

The project could lead to the development of a silicon cerebellum that would allow robots to interact with environmental stimuli in real-time. But this won’t happen overnight. Scientists still need to come to a consensus on a standard working model for the cerebellum and the robotic systems integration will also take time.

The researchers are aiming for a complete understanding of how this region of the brain works. This could one day open the door to better treatments for motor neuron diseases.