ARMTRAK; A Domain for the Unified Study of Natural Language, Planning, and Active Vision.

摘要

ARMTRAK is a micro-world, based on the control of model trains, designed to integrate work in natural language, planning, vision and robotics. The primary advantage of the domain is it provides examples that involve only few objects but that require sophisticated analysis. Because the examples involve few objects, the complex reasoning required is not intractable. On the other hand, more objects can be introduced to study techniques for tractable reasoning. Simple and complex examples in the same domain allow work at different levels to take place simultaneously. As a planning domain, ARMTRAK allows exercising planners in a real time domain about which the planner has only imperfect knowledge. As a domain for natural language research, it allows research into the grounding of language in real situations, and the problem of coordinating the behavior of agents through language. As a domain for active vision research, it is challenging because it requires extracting information whose parameters cannot be completely specified beforehand. Two implementations of ARMTRAK have been developed: a simulation and a version using the Rochester Robot. The simulation allows work on real time planning, and a robot version shows the feasibility of a real working system based on model trains. (kr)