Depth first random walk based on symbolic situated action

  • Authors:

  • Serin Lee;Takashi Kubota;Ichiro Nakatani

  • Affiliations:

  • Department of Electronic Engineering, The University of Tokyo, Tokyo, Japan;Institute of Space and Astronautical Science, Japan Aerospace Exploration Agency, Sagamihara-shi, Kanagawa, Japan;Department of Electronic Engineering, The University of Tokyo, Tokyo, Japan and Institute of Space and Astronautical Science, Japan Aerospace Exploration Agency, Sagamihara-shi, Kanagawa, Japan

  • Venue:
  • ROCOM'06 Proceedings of the 6th WSEAS international conference on Robotics, control and manufacturing technology
  • Year:
  • 2006

Quantified Score

Hi-index 0.00

Visualization

Abstract

The approaches to make an agent decide the proper actions for achieving the goal might be roughly categorized into two groups-the classical planning and situated action system. It is well known that each system has its own strength and weakness with its own application areas. In particular, most of situated action systems do not directly deal with general problems given in propositional logic. However, to build an embodied intelligent agent that can perform complicated tasks as well as navigate, we believe its design should be based on both the situated action approach and symbolic processing, which seem to be antithetic to each other. This paper first briefly mentions a novel action selector to situatedly extract a set of actions, which is likely to help to achieve the goal at the current situation, from the relaxed propositional space. After applying the set of actions, the agent should recognize the new situation for deciding the next proper set of actions. By repeating this procedure, the agent is expected to arrive at the goal state. However, since those actions are derived from the relaxed space in which roughly considers the planning problem, this method can be applied only in the deadlock free domain where fatally wrong decisions cannot be made. To solve the deadlock problems, some of subsequent states after applying an action should be considered to avoid meeting the deadlocks. This paper proposes a novel method to make the agent with the situated action selector solve most of deadlock problems without the help of the conventional planner, which situatedly images the lookahead states. If the agent is caught in a deadlock state during imaginarily walking on the lookahead states, then before meeting the actual deadlock, the agent could avoid meeting it. The experimental results of the proposed approach show that the agent can handle most of deadlock problems, and the quality of the resultant path to the goal is mostly acceptable as well as deriving much faster response time than the classical planning.