Reactive Reasoning and Planning

Michael P. Georgeff, Amy L. LanskyAAAI 1987

Abstract

This paper introduces the Procedural Reasoning System (PRS), a framework designed for autonomous agents operating in dynamic and uncertain environments. By integrating beliefs, desires, and intentions (BDI), PRS enables robots to balance goal-directed behavior with real-time reactivity. The system was successfully demonstrated on the "Flakey" robot, showing its ability to handle complex navigation tasks while responding to emergent events without requiring computationally expensive replanning from scratch.

Introduction

In the mid-1980s, robotic planning was largely dominated by static "plan-then-execute" architectures like STRIPS. These systems required complete plan generation before execution, which often led to failure in real-world environments where conditions change rapidly. The authors address this by proposing a system that avoids overcommitment, interleaving planning with execution to remain responsive to environmental shifts. The goal is to create a robot with "attitude"—one that can maintain long-term objectives while remaining sensitive to immediate environmental cues.

The Procedural Reasoning System (PRS)

The core of the proposed architecture is the Procedural Reasoning System (PRS), which formalizes a robot's internal state through "psychological" attitudes.

The PRS architecture, illustrating the relationship between beliefs, goals, intentions, and the interpreter.

The system consists of four primary components:

  1. The Database (Beliefs): Represents current facts about the world and internal states using first-order logic.
  2. Goals (Desires): Defines desired behaviors or states to be achieved over time, rather than just static end-states.
  3. Knowledge Areas (KAs): Procedural schemas containing an invocation condition and a body of subgoals. These represent the robot's "plans" or expertise and can be triggered reactively by new beliefs or proactively by current goals.
  4. The Interpreter: The central reasoning component that selects and manages KAs on a process stack (Intentions).
An example of a Knowledge Area (KA) for a 'GO-TO' task, showing the procedural logic used to achieve a goal.

A key innovation is Metalevel Reasoning, where KAs can reason about the system's own internal processes. This allows the robot to prioritize tasks or manage computational resources during emergencies, such as deciding whether to continue planning or act immediately.

Experimental Evaluation

The system was evaluated using the "Flakey" robot at SRI and a simulated Space Station environment. Flakey was tasked with navigating a complex layout of rooms and hallways. During these missions, the robot was subjected to simulated "jet malfunctions" (via keyboard input) to test its reactive capabilities.

Results indicated that PRS could successfully interrupt long-term navigation plans to diagnose and address faults, subsequently resuming its original mission. This demonstrated a tight coupling between high-level reasoning and low-level sensing, proving that the robot could maintain alignment in hallways and avoid obstacles while pursuing complex goals.

Conclusion and Future Work

PRS introduced the BDI architecture, which has since become a cornerstone of multi-agent systems. By utilizing "precompiled" procedural knowledge and partial commitment to plans, the system achieves a high degree of robustness in dynamic settings. Future work identified includes improving coordination between multiple PRS instances and refining the knowledge engineering process required to create KAs for increasingly complex real-world geometries.

Summary updated

BibTeX
@inproceedings{georgeff1987reactive,
  author = {Michael P. Georgeff and Amy L. Lansky},
  title = {{Reactive Reasoning and Planning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1987},
}
Figure