Purposive behavior acquisition for a real robot by vision-based reinforcement learning

Abstract

Abstract. This paper presents a method of vision-based reinforcement learning by which a robot learns to shoot a ball into a goal. We discuss several issues in applying the reinforcement learning method to a real robot with vision sensor by which the robot can obtain information about the changes in an environment. First, we construct a state space in terms of size, position, and orientation of a ball and a goal in an image, and an action space is designed in terms of the action commands to be sent to the left and right motors of a mobile robot. This causes a "state-action deviation" problem in constructing the state and action spaces that reflect the outputs from physical sensors and actuators, respectively. To deal with this issue, an action set is constructed in a way that one action consists of a series of the same action primitive which is successively executed until the current state changes. Next, to speed up the learning time, a mechanism of Learning from Easy Missions (or LEM) is implemented. LEM reduces the learning time from exponential to almost linear order in the size of the state space. The results of computer simulations and real robot experiments are given.

Integration of Genetic Programming and Reinforcement Learning for Real Robots

“…The state space was structured based on positions from where the box and the goal area can be seen in the CCD image, as described in [4]. The viewing angle of AIBO CCD is so narrow that the box or the goal area cannot be seen well with only one-directional images, in most cases.…”

“…The "state-action deviation" problem should be taken into account when executing Q-learning with the state constructed from a visual image [4]. This is the problem that optimal actions cannot be achieved due to the dispersion of state transitions because the state composed only of the images remains the same without clearly distinguishing differences in image values.…”

“…The huge amount of learning time required presents a great problem when using a real robot. Accordingly, most studies deal with the problems of receiving an immediate reward from an action as shown in [3], or loading the results learned with a simulator into a real robot as shown in [4,5].…”

Cooperative Behavior Acquisition in a Multiple Mobile Robot Environment by Co-evolution

“…We have selected a simplified soccer game consisting of two or three robots as a testbed for the problem because both competitive and cooperative tasks are involved as stated in RoboCup Initiative [4]. We built an original soccer simulator which models real mobile robots we have been using so far in [1,8,9]. The environment consists of a ball and two goals, and a wall is placed around the field except the two goals.…”

“…Table 1. Although we design these behaviors by hand in this experiments, these primitive behaviors can be acquired by other learning algorithms such as ones in [1,8,9]. …”