Back
to indexJ. K. Rosenblatt
Summary
Describing the behavioral control system evolved at CMU, this paper focuses on the algorithms to control the standard Arbiter components: avoid obstacles, follow road, seek goal, maintain heading, and avoid tip-over. Each behavior is assigned a weight in controlling the vehicle. Behaviors may issue votes at any time. There are several layers of behavioral control: reactive, lookahead behavior, and deliberative behavior. DAMN is a fusion of top down and bottom up algorithms: shared modules have concurrent responsibility for control of the robot. Plans for the robot are used a source of advice, creating a "free flow hierarchy". The arbiter must be capable of satisfying multiple, possibly conflicting goals, which is accomplished by a subsumption heirarchy. As such, the information utilized by each module to make a decision is hidden after the module has made a decision.
Methods
Internally, each module may use its own representation of the world around it. Real time sensors need not communicate with the mapping module to come up with a model of the world satisfying both modules. Instead, only the vote from the module is output to the Arbiter, which then weights the vote and calculates the results together to find the highest-ranking direction to turn. The directions are discrete: hard right, soft right, forward, soft left, hard left. The calculation is based on a normalized weighted sum, to which a smoothing function is applied, and the pixels are interpolated to have finer control than the discrete directions listed above. After this process is complete, the vehicle needs to determine how to apply speed, current turn radius, etc for complete control.
An alternative approach is considered based on the theory of "utility" - each turn of the wheel translates into a certain utility for each module. Utility can be expressed by a Maximum Expected Utility (MEU) measurement, and uncertainty in the system can be reasoned about. Rather than reasoning in discrete directions, utility theory allows a finer degree of control that does not conflict with the current state of the system.
Keywords
behavioral control, hierarchitcal models, utility theory, sensor fusion
Rating
8
Bibtex Entry
@inproceedings{ rosenblatt95damn,
author = "J. K. Rosenblatt",
title = "{DAMN}: {A} Distributed Architecture for Mobile Navigation",
booktitle = "Proc.\ of the {AAAI} Spring Symp. on Lessons Learned from Implememted Software Architectures for Physical Agents",
address = "Stanford, CA",
year = "1995",
url = "citeseer.nj.nec.com/article/rosenblatt97damn.html"
}