2005 Fiscal Year Final Research Report Summary
Intelligent System Controlling Autonomous Mobile Robots with an Intelligent Human-machine Interface and Research on Swarm Intelligence Acquired by Reinforcement Learning
| Project/Area Number |
15500125
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Perception information processing/Intelligent robotics
|
|---|
| Research Institution | Shibaura Institute of Technology (2004-2005)
Kinki University (2003) |
|---|
Principal Investigator |
IGARASHI Harukazu Shibaura Institute of Technology, College of Engieering, Professor, 工学部, 教授 (80288886)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
KUROSE Yoshinobu Kinki University, School of Engineering, Professor, 工学部, 教授 (00043802)
IOI Kiyoshi Kinki University, School of Engineering, Professor, 工学部, 教授 (90288887)
|
|---|
| Project Period (FY) |
2003 – 2005
|
| Keywords | Autonomous Mobile Robots / Reinforcement Learning / Multi-agent System / RoboCup / Pursuit Problem / Path Planning / Behavior Learning |
|---|
| Research Abstract |
In this research project, we constructed a soccer robot system that consists of a host PC, including a frame grabber, five mobile robots, and a video camera set-up over the robot field. Using this robot system, we conducted research on the following four subjects.
The first subject is robust color extraction that does not depend on lighting conditions. Accurate color extraction is necessary so that the robot system can accurately recognize robot markers. We proposed a color recognition method based on a database of threshold values to extract background color (green), the colors of two robot markers (yellow and blue), and the color of a ball (orange). In our experiments, highly accurate recognition rates were observed, even if a deep shadow in a belt-like shape existed in the robot field.
The second subject is the running control of mobile robots with four omniwheels. We used Q-learning to teach the running property to each robot and performed an experiment in which robots moved directly toward a goal. Experimental results showed that Q-learning reduces 50% of the angle deviation from a direct path to the goal.
The third subject is a policy gradient approach to agent learning in a multi-agent system. We applied this approach to such tasks as pursuit problems with obstacles, an inverse problem in a curling-like game, and a collaboration problem involving a kicker and a receiver at free kicks in a soccer simulation game. Our experiments showed the effectiveness of this approach.
The final subject is a portable pointer for positional data acquisition. We developed a portable pointer device and a software system to recognize where the device is pointing.
|
Research Products
(25 results)
