A Fuzzy Controller for Flakey, the Robot

Alessandro Saffiotti, Nicolas Helft, Kurt Konolige, John D. Lowrance, Karen L. Myers, Daniela Musto, Enrique H. Ruspini, Leonard P. WesleyAAAI 1993

Abstract

In the early 1990s, autonomous mobile robots faced a significant challenge in reconciling high-level symbolic planning with low-level reactive control. This paper presents a fuzzy logic-based controller implemented on "Flakey," a mobile robot designed for navigation in real-world office environments. By utilizing fuzzy desirability functions and behavior blending, the system achieves smooth, robust navigation capable of handling sensor noise and dynamic obstacles while remaining faithful to strategic goals. The approach was validated through successful deployment in unmodified office spaces and a second-place finish at the inaugural AAAI International Robotic Competition.

Introduction

Traditional artificial intelligence in the late 20th century often struggled to bridge the gap between abstract reasoning and physical execution. Robots like "Shakey," an earlier predecessor, were often hindered by a rigid adherence to symbolic plans that could not easily adapt to the unpredictability of the real world. When faced with an unexpected obstacle, such systems frequently stalled or failed because their logical frameworks could not reconcile long-term objectives with immediate environmental changes.

This research introduces a novel fuzzy controller for Flakey, a robot designed to navigate the hallways of SRI International. The goal was to create a system that could follow complex plans while remaining inherently reactive—ensuring that the robot could dodge a person or navigate a cluttered corridor without losing sight of its ultimate destination.

The Fuzzy Control Framework

The core innovation of the system lies in its use of fuzzy logic to manage uncertainty and goal prioritization. The architecture is built upon three primary components:

1. Desirability Functions

Unlike binary logic, which might categorize an action as simply "correct" or "incorrect," the Flakey controller uses fuzzy desirability functions. For any given goal (e.g., "reach the end of the hallway"), the system assigns a degree of desirability to a range of possible actions. This allows the robot to evaluate multiple potential paths based on how well they satisfy the current objective under existing environmental constraints.

2. Control Structures and Context

High-level goals are organized into "control structures." Each structure defines a specific behavior, such as "follow-corridor" or "cross-door," and includes a "fuzzy context of applicability." This context allows the robot to determine when a specific behavior is relevant based on its current state and sensor data. For example, as the robot approaches a doorway, the desirability of the "cross-door" behavior naturally increases, triggering a smooth transition between different modes of movement.

3. Behavior Blending

The "secret sauce" of the system is behavior blending. Using fuzzy inferential procedures, the controller can combine multiple desirability functions simultaneously. If Flakey is tasked with reaching a desk but encounters a person in its path, the logic blends the "reach goal" behavior with an "avoid obstacle" behavior. The result is a single, fluid motion that veers around the obstacle while maintaining progress toward the target, avoiding the jerky "stop-and-start" movements typical of earlier robotic systems.

Experimental Validation

Flakey’s performance was evaluated in two demanding settings that tested its ability to handle real-world complexity without the need for high-precision geometric maps.

SRI International Office Environment

The robot was tested in an unmodified office space during peak working hours. This environment featured "normal office activity," including moving people and rearranged furniture. Flakey demonstrated that it could navigate using only sparse topological maps—simple nodes and links annotated with approximate metric information—rather than requiring pixel-perfect environmental data.

AAAI-92 Robotic Competition

The system was put to the ultimate test at the First AAAI International Robotic Competition in 1992. Competing against other leading robotics projects of the era, Flakey placed second overall. The robot was specifically lauded for its "smooth and reliable reactivity." One judge famously noted that Flakey was the "only robot I felt I could sit or lie down in front of," a testament to the safety and predictability afforded by the fuzzy control approach.

Discussion and Insights

The success of Flakey highlighted several key insights for the field of robotics:

  • Bridging the Symbolic-Continuous Divide: The research demonstrated that traditional AI planning can be successfully linked with continuous control systems through fuzzy logic.
  • The Value of Smoothness: By blending behaviors rather than switching them abruptly, the robot achieved a level of natural movement that enhanced both its efficiency and its safety around humans.
  • Robustness to Imprecision: The system proved that perfect sensors are not a prerequisite for autonomous navigation. A controller designed to handle "gray areas" can thrive in environments where data is noisy or incomplete.

Limitations and Future Directions

While the fuzzy controller was highly effective, the authors noted areas for further development. The integration between high-level planners and the fuzzy control structures remains complex, and the paper does not detail the specific computational overhead required for real-time desirability calculations on 1993-era hardware. Additionally, as environmental complexity increases, scaling the system to manage dozens of competing goals simultaneously presents a significant challenge for future research.

Summary updated

BibTeX
@inproceedings{saffiotti1993fuzzy,
  author = {Alessandro Saffiotti and Nicolas Helft and Kurt Konolige and John D. Lowrance and Karen L. Myers and Daniela Musto and Enrique H. Ruspini and Leonard P. Wesley},
  title = {{A Fuzzy Controller for Flakey, the Robot}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1993},
}
Figure