next up previous contents
Next: A Graphical User Up: Mobile Robotics Previous: Software Development Environment

Task-Driven Behavioral Navigation for a Mobile Robot

We are investigating 2D navigational tasks for an autonomous mobile robot possessing sonar sensors and two controllable range sensors mounted on pan-tilt heads (QUADRIS). Experimentation is being conducted within a laboratory and office environment. We believe that our implemented architecture succeeds in overcoming the shortcomings of existing behavioral architectures. The major aspects of our proposed behavioral architecture are as follows: (1) A language lexicon is used to represent spatial information and define task commands. User-specified commands and internal communications are formulated using this lexicon. The chosen lexical subset has been found from the literature to be a minimal spanning basis set for human 2D navigational tasks. The task is either to go to a known location in space or to find a particular object in the environment. (2) An extension of teleo-reactive (TR) trees proposed by Nilsson - called TR+ trees - is used for specifying behavioral control in an asynchronous and concurrent implementation. TR trees are an ordered list of production rules which are continually recompiled during execution time into the equivalent of a hierarchical circuit. A condition is represented by a node while an action is represented by an arc. TR+ trees allow condition and action expressions and present the programmer with the ability to control how and when the expressions are evaluated. The formalism permits the ``personality'' of the robot (i.e., the set of behaviors) to be programmed in a similar fashion to conventional programming (i.e., with parameter passing and binding, hierarchy, and recursion). We have implemented a distributed model of a TR+ interpreter using a tool called PVM. The PVM software system provides a unified framework for developing parallel programs on a collection of heterogeneous computer systems. A graphical programming tool for creating TR+ programs has also been developed. (3) Biologically plausible potential fields are used to perform dynamic path planning. The potential field is modelled as a harmonic function, which inherently does not have any local minima. The path planning is also implemented using a distributed architecture. The implemented framework is such that the computation and execution of the plan are done concurrently. A hierarchical coarse-to-fine strategy is used to guarantee a correct control strategy at the expense of accuracy. A path is recomputed in 3 seconds for a grid of 75 by 75 spatial elements, whereas for coarser levels the recomputation is relatively instantaneous. The implementation of the system utilizes the computer network at the Centre for Intelligent Machines and communicates with the robot via radio and video links. If at least one path exists to a known destination location, the path planning strategy is guaranteed to find a path to that goal. If the destination is unknown, a TR+ program executes an exploratory set of actions appropriate for the sensed environment, context, and task. The higher level reasoning component - COCOLOG - will complement the system by understanding and validating the plausibility of actions and the environmental state.
J. Zelek, M.D. Levine



next up previous contents
Next: A Graphical User Up: Mobile Robotics Previous: Software Development Environment



Thierry Baron
Mon Nov 13 10:43:02 EST 1995