AIR-SOAR: Intelligent Multi-Level Control

Douglas J. Pearson, Randolph M. Jones, John E. LairdAAAI 1993

Abstract

Air-Soar is an autonomous flight agent developed using the Soar cognitive architecture to address the complexities of multi-level control in high-fidelity flight simulation. By decomposing flight tasks into five hierarchical problem spaces, Air-Soar manages simultaneous achievement goals, such as reaching a destination, and homeostatic goals, such as maintaining stability. This paper details the system's ability to handle asynchronous sensing and control, demonstrating robust performance during autonomous missions and resilience against external interference and simulated turbulence.

Introduction

Controlling an aircraft is a "wicked" problem for artificial intelligence. It requires operating in a dynamic, real-time environment where sensing and control are asynchronous, and feedback from actions is often delayed. A primary challenge in autonomous control is the need to balance achievement goals (reaching a specific state, like a new altitude) with homeostatic goals (maintaining a state, like keeping the wings level).

In traditional systems, focusing too much on a high-level task can lead to a loss of control at the low level. Conversely, a purely reactive system might struggle to follow a complex flight plan. Air-Soar aims to bridge this gap by providing a control structure that is sensitive to different "grain sizes" of environmental change, ensuring that high-level mission objectives do not compromise the fundamental stability of the aircraft.

Methods and Architecture

Air-Soar is implemented within Soar, a classic cognitive architecture designed for general problem-solving and learning. The system's intelligence is organized into five distinct problem spaces, each operating at a different level of granularity.

Hierarchical Decomposition

High-level goals, such as "perform a banked turn," are broken down into specific constraints for lower-level spaces. These constraints eventually bottom out in direct "stick commands" sent to the flight simulator. This hierarchy ensures that every high-level intent is translated into actionable, low-level physical movements.

Simultaneous Goal Execution

Unlike systems that switch between tasks sequentially, Air-Soar’s levels are active simultaneously. This allows the agent to perform complex maneuvers, such as climbing and turning at the same time, while continuously monitoring for wing-level stability.

Multi-Modal Reactive Monitoring

Rather than relying on a single, monolithic monitoring loop, each problem space is responsible for noticing when its specific parameters deviate from their constraints. This allows for multi-modal reactivity: a slow heading drift is caught by a high-level space, while a sudden gust of wind is caught by a low-level, high-frequency space. This distributed monitoring makes the system more efficient and naturally reactive to environmental changes.

Experimental Environment

The system utilizes the Silicon Graphics flight simulator, which provides a realistic, high-fidelity environment. The aircraft’s state is updated 20 times per second, providing a continuous stream of real-time telemetry—including throttle, ailerons, and elevator positions—which the agent processes during active flight runs.

Key Insights

Sensitivity at Multiple Grain Sizes

One of the most significant features of Air-Soar is how the hierarchy handles different types of environmental "surprises." If a turn ends and the plane is slightly off-level, the heading will drift slowly; a high-level space eventually notices this and corrects it. However, if a sudden burst of turbulence hits, the low-level "turn rate" space notices the spike immediately and reacts before the high-level heading has even changed significantly.

Implicit Monitoring and Homeostasis

Instead of explicitly checking every parameter against a list of "good" ranges, the problem spaces are designed to notice deviations as they occur within the context of their current goals. By treating stability (homeostasis) as a continuous constraint that interacts with achievement goals, Air-Soar avoids the "tunnel vision" that often plagues autonomous agents during complex tasks.

Results

The capabilities of Air-Soar were demonstrated through a full flight profile, including takeoff, leveling off, following a pre-set flight pattern with multiple turns and altitude changes, and returning to land on or near the runway.

To test reactivity, researchers manually interfered with the flight by moving the control stick while Air-Soar was flying. The agent responded immediately to this simulated turbulence, recovering from forced rolls to re-establish level flight without losing track of its higher-level mission objectives. This demonstrates a robust ability to maintain mission integrity while reacting to unpredictable external forces.

Conclusion and Future Work

Air-Soar demonstrates that a hierarchical approach to cognitive architecture can successfully manage the competing demands of high-level mission planning and low-level reactive control. While this research focuses on the control structure, future work could leverage Soar’s learning capabilities to improve flight performance through experience. Additionally, expanding the environment to include multi-agent interactions would provide further insight into how Air-Soar navigates complex, crowded airspaces.

Summary updated

BibTeX
@inproceedings{pearson1993air,
  author = {Douglas J. Pearson and Randolph M. Jones and John E. Laird},
  title = {{AIR-SOAR: Intelligent Multi-Level Control}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1993},
}
Figure