A Learning Based Vision Guided Robotic Agent Replanning Framework and a Case Study

In this paper, a replanning framework and a case study based on our learning based robotic replanning framework that can handle unexpected events in dynamic worlds are presented. This study presents an original replanning method which uses an alternative based action selection mechanism to select the most efficient action path among possible alternative action paths. The method stores the costs of actions paths from previous executions as an experience and uses it for improving its future action selection decisions. The fact that a continous vision feedback is supplied to the symbolic planning level (highest level of abstraction) is also a contribution of this study since this way, the architecture detects the presence of unexpected events, generates an updated model of the environment and discards or modifies the existing obsolete action path in case of an unexpected event(s).