Introduction
Published in the proceedings of AAAI-96, Murray Shanahan’s work represents a pivotal moment in the "Great Robotics Debate" of the 1990s. It provides a formal logical defense of Cognitive Robotics against the then-dominant "subsumption architecture" championed by Rodney Brooks. Shanahan demonstrates that a mobile robot can build a reliable internal map from noisy, incomplete sensor data using formal logic, bridging the gap between high-level symbolic reasoning and the messy reality of robot hardware.
The Challenge of Noise
In the early 90s, the robotics community was deeply divided. Traditionalists believed robots should build logical models of the world, while critics argued that the "common sense informatic situation"—characterized by noisy sensors and imprecise actuators—made such models impossible. Brooks famously claimed that "the world is its own best model" and that explicit representation was a dead end.
Shanahan addresses this by asking: Can we provide a rigorous, logical account of how a robot can assimilate noisy sensor data into a coherent world model? The difficulty lies in the perceived brittleness of logic. Shanahan’s solution allows logic to "bend" with uncertainty without breaking, treating sensor noise as a form of non-determinism.
Methods: Logic in a Noisy World
The paper utilizes three primary formal tools to handle the informatic situation:
- Circumscriptive Event Calculus (CEC): A many-sorted logic language used to reason about actions and fluents (properties that change over time).
- Abduction: Unlike deduction, abduction reasons from effect to cause. The robot observes a "bump" sensor event and abduces (hypothesizes) the existence and shape of an obstacle that explains that event.
- Non-Determinism as Noise: Instead of complex probabilities, Shanahan models motor noise as a "circle of uncertainty." The robot's location is defined not as a point, but as an ever-expanding region where it might be.
The core innovation is weak abduction. Instead of finding a world model that strictly entails the sensor data, the robot looks for a model that is consistent with the data given the non-deterministic nature of its movement.
Key Insights
- Noise as Non-Determinism: A noisy motor can be modeled by a logical rule stating that after moving a distance $d$, the robot is within a circle of radius $\epsilon \cdot d$.
- Relative vs. Absolute Uncertainty: While absolute position uncertainty grows, relative knowledge (e.g., the shape of a corner) can remain precise.
- Map-Building as Hypothesis Generation: A map is a set of logical hypotheses about objects that best explains the history of sensor readings.
Results and Impact
Shanahan proves that in the absence of noise, this non-deterministic framework is equivalent to standard deterministic abduction. The algorithm was successfully implemented on a "Rug Warrior" robot, which generated logical maps of its environment after exploration. This work successfully demonstrated that noise and incompleteness are not insurmountable barriers to symbolic AI in robotics, preserving a role for representation in intelligent systems.
Summary updated