|
System /
ATLANTISA Three-Layer Architecture for Navigating Through Intricate SituationsA terribly named system which implements a Three Level architecture, with ALFA fulfilling the reactive layer and RAP as the sequencer. A generic General Planner handles deliberation. Described in Papers#Gat92Integrating. Developed by Erann Gat and is one of the predecessors to 3T. ActivitiesSeems like the most important thing here is the idea of continuous actions, or activities. These are the base abstraction for what the robot can do; the building blocks for behaviors. They are described in ALFA, and then arranged by RAP. I think. Or something. Gat differentiates this system from traditional AI systems, which rely on state-action models. Actions are atomic, and result in a change from one state to the next. This is the basis for the models used in Remote Agent (which is a Model Based approach). The ease of analysis of this model is powerful, but it does not model the environment of most robots well (space, the domain of most of the Remote Agent work, is an exception. At least, it is close enough.). But activities acknowledges that actions should be continuous. The activities he describes are hierarchical, so that when a decision sets off an activity, that activity may itself activate other activities. This stops when primitive activities are reach, which are basic sensor-motor processes (and reactive). Gat cites the following papers for the continuous action model (I have not investigated):
|