| 102B4 |
|---|
| Acquisition of suitable actions in a real mobile robot with a CCD camera by Direct-Vision-Based Reinforcement Learning |
| AUTHORS |
| Masaru IIDA, Masanori SUGISAKA, Katsunari SHIBATA (Oita university) |
| ABSTRACT |
| In this paper, it is confirmed that Direct-Vision-Based Reinforcement Learning (RL) is effective on-line action learning of a real mobile robot with a large number of visual signals from a CCD camera. In Direct-Vision-Based RL, raw visual sensory signals are put into a layered neural network directly, and the neural network is trained by Back Propagation using the supervised signal generated based on RL. It has been shown that the learning is fast and stable, and the robot could obtain a global representation in the hidden layer by integrating the visual sensor signals, each of which represents local information in some simulations. It was formerly shown that the real mobile robot with a monochrome visual sensor consisting of 64 cells could obtain appropriate reaching actions to the target object through Direct-Vision-Based RL from scratch.In this paper, a CCD camera was implemented to the robot, and the 64*24=1536 visual signals were the input of the neural network. Although it was afraid that the neural network could not learn appropriate reaching actions to a target object because of a large number of the input signals, it is shown that the robot could obtain appropriate reaching actions through the learning from scratch without any advance knowledge in "going to a target" task. |
Back to Program
Time Table
Top of This Site
Annual Conference 2002