Abstract
An autonomous mobile robot must be capable of rationally selecting its tasks in order to achieve some state or goal. Rule-based systems encode system knowledge into a set of rules that guide the robot. The dynamic nature of the world is seldom consistent enough to support the rigidity of rules. A new decision maker is needed that expresses the problem as intuitively as rule-based systems with flexibility, extensibility and generalizability.
The constraint-based methods can be used to model the problem of selecting and configuring tasks for mobile robots. Constraint Satisfaction Problems (CSPs) provides a framework in which multiple conflicting constraints upon the robot can be resolved in such a way that the robot not only performs correctly, but also meets or exceeds its performance requirements. Constraints also provide the intuitive means of specifying the decision model without the rigidity of rules. Performance objectives are incorporated into the constraint model so that the decision-making system is capable of rationally guiding the robot through actions that best meet its current needs and goals.
In this dissertation, a framework and decision maker for robot task selection and configuration is developed. Tasks under this framework are modeled as constraint satisfaction problems. A common software interface was used in order to support the uniform composition of the CSP models. The solution to the CSP provides the selection of a task and its configuration. A utility function is used to select a single solution, if multiple solutions are generated. The framework is demonstrated for three unique robot scenarios: a delivery robot, a polar robot, and an urban search and rescue robot.
The delivery robot scenario models tasks for a mobile robot responsible for delivery of items within an office environment. It provides an initial proof-of-concept. Constraint models are constructed for a charge task, a pickup item task, and a deliver item task. The solution to the CSP configures the task execution by specifying the satisfying speed and path for the robot to follow. The performance of the system given the task load, which is directly proportional to the size of the CSP model, is evaluated. Results show a significant increase in computation time when the number of simultaneous delivery requests grow beyond five tasks; however, the delivery failure rate remained lower than a traditional rule-based approach as the load increased.
The polar robot scenario models a autonomous mobile robot for remote sensing of Polar Regions The simulation of the polar robot includes a number of remote sensing instruments, including a synthetic aperture radar (SAR), accumulation radar, gravimeter, magnetometer, and IR spectrometer. The robot is also equipped with a solar and a wind generator. The challenge was to balance robot survival and data collection over a full Antarctic year. Constraint models for each instrument, generator, and task are implemented and evaluated versus a rule-based system. The constraint-based system produced significantly lower failure rates, 70% or lower, versus a near 100% failure rate for the rule-based system. The mean survival time using the constraint-based decision maker is greater than 250 days; and the rule-based systems mean survival is less than 200 days. The mean mission completeness of the constraint-based system is significantly greater than the rule-based system, at a 95% confidence level, for four out of five experimental configurations of the polar scenario.
The urban search and rescue scenario (USAR) models a mobile robot for the mapping and exploration of collapsed buildings to locate victims and hazards. The robot and its environment are simulated based on works on robot-assisted search and rescue. Task models are constructed for searching, reporting results to the rescue party, obtaining repairs, and charging. Eight unique experiment configurations are developed with varying victim injury levels,number of blocked locations, and topologies (hospital and hotel). The constraint-based system performed statistically better than the rule-based system for two out of eight configurations for mean victims rescued; two out of eight for hazards localized; four out of eight for mean number of collisions; and four out of eight for locations mapped. The constraint-based framework meets or exceeds the performance of the rule-based system.
The new decision framework is capable of guiding a variety of mobile robots through rational decisions for task selection and configuration. This is demonstrated in this dissertation for three different scenarios. The framework is flexible to changes in the environment, mission, or tasks. It is also extensible as constraint models can be extended to develop models for new tasks. This work demonstrates that constraint-based decision making is a viable apporach to robot task selection and configuration, and performs better than rule-based systems over a variety of applications.
Original language | American English |
---|---|
Qualification | Ph.D. |
State | Published - Jul 2007 |
Disciplines
- Artificial Intelligence and Robotics