1. Introduction

　The purpose of this study is to develop a robotic endoscope that is low invasive, easy to operate and locomotesfrom the rectum to the appendix in human body.

2. Background

　Incidence of colorectal cancer ranks fourth in men and third in women with over 1 million new cases occurring every year worldwide. Five-year survival estimates are around 50%. However, the patients can recover fully if the colorectal cancers are detected early. Thus, early detection is one of the most important factors to treat colorectal cancers. Colonoscopic examination is one of the most reliable methods to detect colon cancers of early stage.

3. Mechanism

3.1. Rotary inertia type

The rotary inertia type has a small DC motor with the flywheel and it is covered with spiral fin. The flywheel is accelerated slowly and it is rapidly decelerated when the flywheel reaches the highest velocity. So, the main body rotates by the inertia force. As a result, the robot is moved forward by spiral shape fin. This model's advantage is big energy can be conserved and big inertia force can be outputed continuously.

3.2. Reverse screw type

The reverse screw type has two units which has each clockwise screw fin and anticlockwise screw fin. And, these units are rotated by the DC motor. The opposite direction of the screw and the opposite direction of rotation generate one direction thrust force. This model's advantage is continuous and smooth locomoting. 　We are developing this method now.

3.2.2 WQE-2（Waseda Q:Kyudai Endoscope No.2）

3.2.2.1 WQE-3（Waseda Q:Kyudai Endoscope No.3）

3.2.2.2 Experiment for evaluation of WQE-3

3.2.3 Optimization of fin lead

4. Reinforcement Learning

　The best method of the robot control is trial-and-error when the robot locomote in the colon. Because, locomoting of the robot influences the colon environment and it is impossible to model . Then, the endoscope robot is controlled with the reinforcement learning. The feature of the reinforcement learning is as follows.

• Learn while interacting with environment
• Learn automatically according to an obtained goal
• Adapt uncertain environment

　It studied by using an actor critic method in the reinforcement learning. The feature is as follows.

• Drive continuously according to probability
• Program easily

　They are movies of robot locomoting with the reinforcement learning and without the reinforcement learning. Click the following pictures.
(※The image of the internal organs of swine flows.)

　This graph is relation between the frequency of the reinforcement learning and the movement distance in dead swine.

　This research is done under instruction of Dr. Tanoue, Dr. Ieiri, Dr. Konishi, Dr. Uemura, Dr. Ohdaira, Dr. Tomikawa and Dr. Hashizume of a Kyushu University Hospital Center for the Integration of Advanced Medicine and Innovative Technology.
We would like to express our thanks to all co-researchers ,Kuraray Co, Ltd., Okino Industries, Ltd., and SolidWorks Japan K. K. for helping us to develop the endoscope robot.