Figure 3 - uploaded by James Brusey
Content may be subject to copyright.
Source publication
A central problem in autonomous robotics is how to design programs that determine what the robot should do next. Behaviour-based control is a popular paradigm, but current approaches to behaviour design typically involve hand-coded behaviours. The aim of this work is to explore the use of reinforcement learning to develop autonomous robot behaviour...
Citations
... In terms of reactive archi-tectures, we may draw a distinction between purely reactive and behavior-based systems. Purely reactive systems, which are also called reflexive systems, can be represented by a set of mappings between input and output, for example, from sensors to actuators [2]. They have no internal memory and, instead, use the state of the environment as a form of memory. ...
The architecture of a modular behavioral agent (MBA) with learning ability for hardware realization is proposed to implement a multilevel behavioral robot such that it can autonomously complete a complex task. The architecture is composed of similar modules, primitive actions, and composition behaviors. These modules, which are derived from a basic template, are capable of learning and cooperating to cope with a variety of tasks. The infrastructure of a template embeds a reinforcement learning mechanism with an adaptable receptive module (ARM)-based critic-actor model. Each template executes one specified behavior and also cooperates with other templates to form a more dexterous composed behavior. In other words, the composed behavior is constructed by several primitives with similar modular architecture. The learning and cooperation abilities in the modules are based on a reinforcement learning technique, which is based on a critic-actor model. The proposed architecture is implemented in a field-programmable gate array (FPGA) chip with a CPU core such that the computing device can fully utilize the merits of parallel processing of neural networks in the ARM scheme. The study is demonstrated on a mobile robot for goal-seeking, cruise , and safety ensurance tasks in an unstructured environment with obstacles such as walls and blocks. The results show that this robot with the modular architecture can perform well in unstructured environments.
... Although learning has been present since the creation of this league (see for example the use of Genetic Programming in [Luke et al., 1998]), the focus has been more on cooperation of well-defined simulated agents (one's own team), rather than learning the behaviour of the opponent team. When learning is involved, it is most of the time to learn basic behaviours (controlling the ball, passing the ball) or learning how to combine pre-defined basic behaviours to form a team strategy (see [Salustowicz et al., 1998] or [Brusey, 2002] and, for an account of layered learning, [Stone and Veloso, 1998]). ...
The work presented in this paper aims at combining fuzzy function approximation and reinforcement learning in order to create robotic soccer agents that are able to coordinate their behaviours locally and socially while learning from experience. This simultaneous coordination and learning ability can play a crucial role in improving the behaviour usage of robotic soccer agents. To achieve this goal, a fuzzy reinforcement learning technique for a single agent is first examined and then this technique is applied to multiple agents. The conducted experiments through a soccer simulation system show that the performance of robot scoring speed is improved using the proposed approach.
