Overview

The combination of SRI's fuzzy control, local perceptual space, and procedural reasoning technologies provides for flexible, robust control of an autonomous vehicle operating within an (unmodified) office environment. Purposeful behaviors (e.g., going to a specific location) are dynamically blended with reactive behaviors (e.g., avoideing a moving obstacle) to provide smooth and compliant control of the vehicle as it navigates according to a strategic plan.

SRI's work in combining sensory information with map-based information has led to the development of the local perceptual space (LPS), an egocentric view of local space where information from all sensors and interpretation routines is posted. Here, information from different sensor modalities, and different levels of abstraction and complexity, freely mixes. As new information is posted, different routines respond by processing it and posting new results; this hierarchical nature of interpretation allows time-critical routines to respond quickly, while more complex routines take more time.

Domain

The problem is to control Flakey, SRI's autonomous robot, as it navigates through an office environment according to a multistep strategic plan, an approximate map of the area, and the information provided by several on-board sensors. The vehicle must be capable of executing the plan despite the presence of unanticipated objects, moving obstacles (e.g., people), and noisy sensors.

See videos of Flakey in action.

Application

The LPS is used as the repository for the results from a collection of interpretation routines on the robot. This includes raw distance readings, extracted surfaces (e.g., walls, obstacles), objects from the map (e.g., specific corridors), and abstract control points from the strategic plan (e.g., positions to obtain).

At any time, several active behaviors are responding to the information in the LPS. Some of these are purposeful (e.g., following a corridor to reach a target location), while others are for contingencies (e.g., avoiding obstacles). Each behavior is defined by a set of fuzzy rules. A rule's level of activation corresponds to the degree that its precondition matches the LPS; when the precondition of a rule holds, then its action is very desirable; when the precondition partially matches, the action is less desirable. By blending the actions of all the rules, in the currently active behaviors, according to their levels of activation, an "optimal" action is derived. Since the conditions in the environment typically change slowly, the repeated application of the same fuzzy rules smoothly blends actions over time. For example, as a person approaches, the robot gradually veers away, and then returns to its original course once the person has passed.

At a higher level, PRS-Lite, a version of SRI's PRS-CL with a 100 milliseconds cycle time, provides real-time supervisory control for the robot. It does so by following a strategic plan that states the sequence of goals that the robot needs to satisfy. This plan may have been automatically produced by a generative planner or directly specified by the user. PRS-Lite utilizes a library of predefined procedures to reduce these high-level goals into subsequences of lower-level goals; at the lowest level, goals are satisfied by activating and deactivating behaviors according to the refined plan and the situation-dependent information in the LPS.

Performance

Our in-house experiments have shown that this methodology leads to robust, flexible controllers that reliably attain user-specified goals while smoothly reacting to unexpected events. Flakey freely roams our offices, coexisting with our staff. Its presence is nonthreatening, and it requires no special accommodations by those around it. The utility of this architecture was apparent in the first two AAAI mobile robot competitions. In the first (1992), Flakey was the only entry capable of successfully performing in the object location and mapping event without requiring environmental modifications. Flakey's multimodal sensor interpretation proved to be very effective, with the sonar routines suggesting candidate objects that were verified by the structured light routines. Flakey received special recognition for its smooth and compliant behavior in the presence of people as exemplified by one judge's comment: ``Only robot I felt I could sit or lie in front of.'' In the 1993 contest, additional visual recognition of landmarks and visual stereo processing for obstacles were incorporated. Flakey finished in the top category in the two events it entered.

Collaborators

Alessandro Saffiotti, Université Libre de Bruxelles

Kurt Konolige <konolige@ai.sri.com>
Karen L. Myers, <myers@ai.sri.com>
Enrique H. Ruspini, <ruspini@ai.sri.com>
Artificial Intelligence Center
SRI International

---

PRS-CL and PRS-Lite are trademarks of SRI International.

Copyright © 1995 SRI International, 333 Ravenswood Ave., Menlo Park, CA 94025 USA. All rights reserved.