Abstract
This paper presents the Operator Compiler, a system designed to bridge the gap between Qualitative Process Theory (QPT) and automated planning. By automatically translating physical models into planning operators for the temporal planner TPLAN, the system enables AI agents to reason about physical manipulations—such as fluid transfer or thermal changes—without the need for manually encoded planning operators. The approach ensures that an agent's plans are physically grounded in a formal model of the world.
Introduction
In the late 1980s, the field of Qualitative Physics (QP) excelled at predicting physical behavior, such as diagnosing faults in a circuit or predicting when water would boil. However, these systems were largely "hands-off," lacking the ability to formulate plans for an agent to intervene in the world. Conversely, traditional AI planners were effective at reaching goals but struggled with the continuous, process-heavy nature of the physical environment. To address this, the author introduces a method to automatically derive planning "manuals" from formal physical descriptions.
Methodology
The system integrates two primary components: the Qualitative Process Engine (QPE), based on Forbus’s Qualitative Process Theory, and TPLAN, a temporal interval-based planner based on James Allen’s logic. The Operator Compiler acts as the bridge between these two, performing a three-step mapping process:
- Entity Mapping: QPT entity definitions (e.g., the properties that define a "container") are converted into TPLAN rules.
- Process Mapping: Physical processes, such as "Liquid Flow," are transformed into planning operators. The compiler identifies necessary preconditions (e.g., a valve must be open) and the resulting effects (e.g., the amount of liquid increases).
- Causal Chaining: The compiler utilizes "qualitative proportionalities" (QPROPs) to allow the planner to back-chain from a goal. If a goal requires increasing a water level, the compiler identifies that this requires increasing the amount of liquid, which in turn requires initiating a flow process.
Temporal Logic Foundations
Because physical processes are not instantaneous, the system relies on Allen’s interval logic to reason about events that occur over time. This allows the planner to handle overlapping events and persistent states.
As shown in the table above, TPLAN uses a set of primitive relations—such as meets (:M), overlaps (:O), and during (:D)—to define how different intervals of time relate to one another. This is crucial for physical domains where, for example, a valve must remain open during the entire interval that water is flowing.
Evaluation: The Kitchen Domain
The system was evaluated using a simulated "Kitchen Domain" featuring faucets, sinks, and various containers. The goal was to test if the planner could coordinate agent actions (like moving a pot) with physical processes (like filling it).
In a typical test case, the system was tasked with increasing the amount of water in a specific pot. The compiler successfully generated a plan that:
- Identified the need for a Liquid Flow process.
- Determined the pot must be positioned under a faucet.
- Invoked a rule for the agent to move the pot.
- Identified that the valve must be opened to trigger the flow.
Discussion and Limitations
While the Operator Compiler successfully automates the creation of planning operators, it faces challenges with physical ambiguity. The system currently assumes that any positive influence on a quantity will result in an increase, failing to account for competing influences (e.g., a sink filling and draining simultaneously). Furthermore, the generated rules are relatively static and may struggle with highly context-dependent environmental changes. Future work is required to handle more complex sets of individuals and more nuanced causal reasoning in ambiguous scenarios.
Conclusion
By deriving planning operators directly from a theory of physics, the Operator Compiler ensures that an AI agent's actions are consistent with the physical laws of its environment. This integration of Qualitative Process Theory and temporal planning represents a significant step toward more robust and physically-aware autonomous agents.
Summary updated