A Dynamical Systems Approach for Robot Learning Problem
スポンサーリンク
概要
- 論文の詳細を見る
Recently, behavior-based robotics has attracted large attentions. Most of behavior-based robots do not involve with complex internal computations, such as planning or inference, instead those studies put more emphasis on "reactive-type" behaviors enabled by simple sensory-motor mapping functions. Although their simple behavior mechanisms resulted in many successes in the real world examples, too much emphasis of this philosophy might restrict the robot's progress in emulating the equivalent cognitive abilities of animals or humans. We consider that the problems of cognition for robots start when they try to acquire certain forms of descriptions of the world. The robots can be regarded as "cognitive" when they can utilize some functions by which they mentally simulate or plan their behavioral sequences using such descriptions.It has been traditionally believed that the mental processes, as like planning or inference, are best represented by symbolic processes; since symbol systems consisting of arbitrary shape of tokens are characterized by their "combinatory power" in representing and manipulating knowledge. For example, one may describe possible interactions between a robot and its environment by a graph of a state-action chain (which could describe combinatory sequences of the robot actions).There have been many studies in this context such as situated automata theory by Rosenschein and qualitative spatial reasoning by Kuipers and Mataric in the navigation domain using the landmark-based graph representation. However, when one tries such symbolic approach, one faces so-called Symbol Gronndmg Problem as discussed by ilarnad. The problem says that there are fatal gaps between what symbols represent and what the real worlds are. Since symbol systems consisting of arbitrary shape of tokens do not share the same metric space with the physical world based on the real number systems, such symbols may not be grounded. In this paper, we try to investigate this problem from the view point of the dynamical systems approach with speculating that the real number systems best represent the mental activities of robots. We can expect that chaos of well-structured may serve as a source of mental or cognitive activities of robots; as have been shown that symbolic dynamics by chaos exhibits a certain linguistic complexity. We further speculate that if equivalences of symbolic systems (as analogical models) can be obtained, as self-organized in the neural dynamics, through direct interactions with the physical environment, such "symbol systems" in terms of real number dynamical function can be naturally grounded.We describe briefly our ideas in the domain of navigation problem and the experiments using a real mobile robot.
- 社団法人電子情報通信学会の論文
- 1996-03-11
著者
関連論文
- A Dynamical Systems Approach for Robot Learning Problem
- Dynamical Systems Approach in Learnable Autonomous Robots
- Evolving Planning Behaviors in an Artificial Ecosystem