Researchers Revamp Asimo Robot's BMI Technology

Oct 22, 2010
Tomonori Shindou, Nikkei Electronics
The magnetoencephalograph (MEG) used for the measurement of a brain
The magnetoencephalograph (MEG) used for the measurement of a brain
[ If it clicks, the expanded picture will open ]

A Japanese research group developed a BMI (brain machine interface) technology that can continuously estimate the movements of the user's finger and reproduce them by noninvasively measuring the person's brain activity.

The group consists of members of Japan's National Institute of Information and Communications Technology (NICT) and Advanced Telecommunications Research Institute International (ATR).

Existing noninvasive BMI technologies estimate right answers by using several patterns of movements registered in advance. But, with the new technology, it is possible to continuously estimate the two-dimensional coordinates of a fingertip (within a range of 20cm) at a frame rate of 50fps (20ms), which is as high as that of a movie.

The spatial accuracy is 14.7mm. The research group expects that the technology will be used for the remote control of a robot and so forth (See related movie).

In March 2009, ATR announced a technology to control the "Asimo," Honda Motor Co Ltd's humanoid robot, with a BMI technology (See related article). And the latest BMI technology was developed by making improvements to it. While the BMI technology for the Asimo can recognize only four pre-registered patterns of movements (category identification), the new technology can estimate the coordinates of a finger tip, which is a continuous parameter.

"Because the latest BMI technology can smoothly reconfigure rapid movements, it gives the user a feeling of controlling a robot by himself," said Hiroshi Imamizu, Biological ICT Group, NICT.

Combining several brain measurement methods

The BMI technology for the Asimo is used by combining an electroencephalograph (EEG) with a high time resolution and a near-infrared brain measurement device (NIRS, optical topography) with a high spatial resolution in a mutually complementary manner. And the researchers employed the same approach of combining multiple measurement devices for the new technology.

Specifically, they combined a magnetoencephalograph (MEG) with a high time resolution and an fMRI (functional Magnetic Resonance Imaging) with a high spatial resolution. Both of the devices are large and expensive. So, they intend to enable to use the latest technology with a combination of an EEG and a NIRS.

Because brain activity differs from person to person, it is necessary for a BMI system to go through a learning process before using the new BMI technology. When the user is moving the finger tip in eight directions, the brain activity of the person is measured individually by the MEG and the fMRI. And the system learns from the data collected by measuring 200 rounds of fingertip movements per user.

(Continue to the next page)

It takes about a day to use the data and establish a model used for estimating the movements of the user's fingertip. However, after the learning process, the system can estimate the movements at high speeds.

This time, the researchers operated the system as an offline system, but they are now researching on a system that operates in real time. In that case, the delay time is about 0.5 seconds. Most of the delay time is the time it takes for an enormous quantity of the MEG's time-series measurement data to be transferred among multiple computers.

Estimation conducted in 2 steps; sparse logistic regression used in latter part

The estimation of the movements of a fingertip is conducted in two steps. First, based on the data collected by the MEG, the current is estimated at 1,500 current sources that are virtually and evenly arranged on the surface of the cerebral cortex. Then, a dimensionality reduction technique called "sparse estimation (SLR: sparse logistic regression)" is used to extract only the current sources that are related to the movements of the fingertip.

As a result, 200 current sources, including pyramidal area, parietal association area and somatosensory area, that are related to the movements are automatically chosen. The learning process for establishing an SLR model takes several hours to complete. The SLR is a dimensionality reduction technique developed by ATR, and it is also used for the BMI technology for the Asimo.

The MEG is equipped with 400 channels of magnetic field sensors, but each sensor is affected by mingling magnetic fields generated in various regions of the brain. The number of the channels used for the measurement (400) is smaller than the number of unknown current sources (1,500). So, it is a so-called ill-posed problem.

To solve the problem, the researchers utilized the data collected by the fMRI. Specifically, an "inverse filter" is established based on the fMRI's data for restoring 1,500 current sources in the brain by using the MEG's data. To establish the filter, they used an fMRI having a high spatial resolution. And, when the BMI is actually used, an MEG that can output data at a speed as high as 1kHz is used. For the learning of the inverse filter, the "variational Bayes approach" was employed.

Removing influence of electric fields generated by myoelectric signals

An MEG is a device that measures the weak electric fields generated by the spike electricity of brain neurons. Therefore, it also measures the electric fields generated by myoelectric signals outside the brain as in the case of an EEG. For example, they are electric fields generated by heartbeat and eye movements.

(Continue to the next page)

"The electric field generated by heartbeat is much stronger than that generated by a brain," said Masaaki Sato, director of the Brain Information Communication Research Laboratory Group of ATR.

To remove such artifacts generated by heartbeat, eye movements and so forth, about 10 current sources are allocated in eye balls, heart and other body parts than brain in the process of estimating current sources by using the inverse filter.

In the learning process, data on the electric fields generated by eye movements, etc are measured by sensors, and they are input as "teaching signals." But, when the new BMI technology is being used, it is possible to remove such artifacts without using those sensors.

To use the latest BMI technology, the user has to actually move his/her finger. But the researchers aim to develop a BMI technology that can be used just by imagining a movement. For the future, they will consider combining an EEG and an NIRS.