Abstract
This paper presents a framework designed to bridge the divide between formal logic-based planning and reactive robot control. By defining "correct" reactive rules as the initial step of a minimal plan derived from a formal theory of action, the authors propose a system that maintains theoretical grounding while remaining robust enough to navigate dynamic and unpredictable environments. The approach was implemented on a B-14 mobile robot for the AAAI 96 robot contest.
Introduction
In the mid-1990s, AI robotics faced a significant schism. On one side was the traditional deliberative approach, exemplified by systems like STRIPS, which relied on high-level reasoning to observe the world and generate complete plans. While logically sound, these systems often failed in real-world applications because planning was computationally expensive, and the environment frequently changed before a plan could be fully executed.
On the other side was the "situated activity" or reactive robotics paradigm. These robots bypassed complex reasoning in favor of immediate reactions to sensor inputs via pre-defined rules. While fast and successful in dynamic settings, these systems often lacked a formal theoretical basis, making it difficult to prove they would achieve specific goals or to reason formally about their behavior. This paper addresses this gap by asking: How can formal theories of action be used to design reactive rules that are guaranteed to be "correct"?
Methodology and the Action Language $\mathcal{L}^1$
The authors introduce a methodology based on a formal action language called $\mathcal{L}^1$. Unlike its predecessors, $\mathcal{L}^1$ is specifically designed to specify dynamic worlds. The core of the proposed architecture is the Reactive Rule:
if f1, ..., fn then a
In this structure, $f$ represents "fluents" (conditions derived from sensor data) and $a$ represents the action to be taken. The primary innovation of this work lies in the definition of correctness. The authors assert that a reactive rule is correct with respect to a goal if, for any situation matching the "if" conditions, the action $a$ is the first action in a minimal plan to reach that goal.
Instead of committing to a long-term plan, the robot follows a continuous loop:
- Sense the environment.
- Identify a rule where the conditions are satisfied.
- Execute only that single action.
- Immediately re-sense and repeat the process.
Implementation and Key Insights
The work was developed for the AAAI 96 robot contest using a B-14 mobile robot from Real World Interface (RWI). The implementation highlights several critical insights into the nature of robotic control:
- The Power of the "First Action": By only executing the first action of a minimal plan and then re-sensing, the robot avoids the brittleness of long-term commitment. This allows the system to handle sensor errors, unpredictable obstacles, and imperfect control. If a sensor provides a wrong reading or an action fails, the subsequent sensing cycle will naturally trigger a corrective rule.
- Formalizing Reactivity: The paper provides a mathematical bridge that allows developers to write reactive rules that are logically derived from a formal theory of action rather than relying solely on heuristics.
- Temporary Perfection: The authors acknowledge that $\mathcal{L}^1$ assumes perfect sensors and control. However, they argue this is a functional abstraction because the robot only relies on these assumptions for the duration of a single action. The constant sensing loop compensates for the inherent imperfections of the real world over time.
Results and Limitations
As a description of an approach for an upcoming competition, the paper focuses on qualitative framework demonstration rather than quantitative speed or success metrics. The authors successfully demonstrate a system where reactive rules can be formally verified against a goal, and the robot can operate in real-time by avoiding the traditional "plan-then-execute" bottleneck.
However, limitations remain. The underlying theory still assumes that action effects are known and sensors are correct at the moment of sensing. While the feedback loop mitigates minor errors, a catastrophic sensor failure could still lead to incorrect action selection. Additionally, the authors note that control modules must be designed to prevent rule conflicts (where multiple rules are satisfied simultaneously), a challenge that may increase in complexity as environments become more sophisticated. At the time of publication, the authors were continuing to refine the integration of high-level reasoning and low-level control for the B-14 platform.
Summary updated