Ken Yamane Laboratory

Our laboratory is conducting research aimed at elucidating the computational aspects of brain information processing and engineering higher-order functions such as inference. Another theme is to demonstrate the usefulness of brain-inspired information processing systems by applying brain-inspired information processing to real-world problems, thereby overcoming the limitations of classical artificial intelligence's information processing capabilities. In particular, we are developing human assistance technologies.
Each member of this laboratory is conducting research on a theme in the fields of neural networks, robots, and artificial intelligence. Through their graduation research, fourth-year undergraduate students learn how to conduct research, write papers, and give presentations. Graduate School further explore their own themes, interact with other researchers, and publish their findings in papers and at academic conferences. They are also working with first- to third-year undergraduate students as research interns to develop an autonomous outdoor mobile robot.

Brain type inference using recurrent neural network
Classic artificial intelligence (classical AI) internally expresses external world information as a symbol and realizes information processing such as inference by manipulating the symbol. However, there are two unsolved problems (symbol grounding problem and frame problem) in classical AI. Therefore, its information processing capacity is limited.
On the other hand, we are focusing on information processing in the brain. In the brain, information is expressed in a distributed manner as activity patterns of many nerve cells (neurons), and the patterns are dynamically converted according to the autonomous dynamics created by the network of neurons. Using these as hints, I am studying brain type inference methods using recurrent neural networks.

Hand / arm motion estimation from surface myoelectric potential signals using brain-type information processing system
We are aiming to develop human support technology that supports human cognitive and motor functions. For example, consider replacing the functions of the hands and arms with a robot arm. To do so, it is necessary to quickly estimate a person's movement intentions and plans. However, meaningful "movements" and "behaviors" that consist of a combination of several movements are very difficult to handle because they change both spatially and temporally.
In this regard, we propose a system that estimates the behavior of the hand and arm from the myoelectric potential signal that can be obtained from the skin surface of the arm, using the orbital attractor model, which is a type of recurrent neural net with continuous time dynamics. I am. Since this system can flexibly estimate complex behavior, it has great potential for application.

Development of outdoor autonomous mobile robot Progress-i
We are developing a robot that can safely and reliably judge and travel to the destination in an outdoor environment prepared for humans. Since it is outdoors, the range of activity is wide, and unexpected things such as pedestrians popping out, tires slipping, and sensors not functioning well due to the weather often occur. It is required to accurately estimate the position of the robot itself and move wisely while recognizing the environment flexibly.
This robot is planned to be used as part of a system that automatically inspects the equipment of a large-scale solar power generation facility (mega solar) in the future. A system that patrols tens of thousands of solar panels one by one, takes pictures from the back side of the panel with a thermal image camera, sends the data to the analysis server, and discovers places (hot spots) that have local heat. is.