Abstract
This paper addresses the historical divide between high-level logical reasoning and low-level reactive robot control. It proposes a framework where reactive behaviors are specified as acyclic logic programs. By modeling the robot as a transduction from sensor history to actuator outputs and deriving a compact state representation, the author provides a mathematically rigorous method to unify robot controllers and physical environments in a single language.
Introduction
In artificial intelligence, a long-standing gap exists between high-level reasoning systems (such as STRIPS) and low-level reactive controllers. While logic-based systems excel at abstract planning, they are often considered too slow for real-time movement. Conversely, reactive control loops lack formal semantics for complex reasoning. David Poole challenges this dichotomy by asking if logic programs can specify the actual control laws of a robot.
Methodology: Logic as Transduction
The core of the proposed framework is the use of Acyclic Logic Programs with a temporal structure. The robot is modeled as a "transduction"—a function mapping a history of sensor inputs to a history of actuator outputs.
To make this computationally feasible, the system identifies "recallable fluents." Instead of re-calculating the entire sensor history at every time step, the system uses a was predicate to access the most recent past value. This effectively derives a "state" from the logic without requiring it to be defined as a primitive concept.
The framework also enables the implementation of numerical methods, such as Euler integration, directly within logic. This allows the same language to define both the robot's controller and the environment's physics, facilitating integrated simulation as shown in the block diagram above.
Evaluation and Results
The approach was evaluated using a simulated nonholonomic robot navigating a maze. The robot must use its sensors to detect walls and steer toward a goal while adhering to steering constraints.
Key results from the simulation include:
- Real-Time Performance: The system achieved over 20 steps per second on mid-90s hardware (a 68040 processor), demonstrating that logic-based control is not inherently slow when state is maintained efficiently.
- Expressiveness: Complex behaviors, including "bang-bang" control and obstacle avoidance, were represented in just a few dozen lines of logical axioms.
- Hybrid System Modeling: The logic successfully derived discrete events (e.g., hitting a wall) from continuous-time movement equations, proving its utility for hybrid systems.
Discussion
The paper suggests that causality in physics is equivalent to acyclicity in logic programs. By maintaining this property, the programs remain stable and well-defined. Furthermore, the logic-based nature of the controller provides a clean interface for handling uncertainty through Probabilistic Horn Abduction, allowing for reasoning about noisy sensors and unreliable actuators.
Future work includes hardware compilation into FPGA/VLSI and the development of adaptive sampling techniques, where the logic program determines the optimal frequency for sensor readings based on the environmental context.
Summary updated