Abstract
Autonomous agents operating in real-world environments must navigate complexities such as continuous time and concurrent events. Traditional AI planning often relies on discrete state transitions, failing to account for "natural actions"—environmental events like tides that occur independently of the agent's will. This paper presents an extension to the FLUX language, integrating Fluent Calculus with Constraint Logic Programming (CLP). By employing process fluents and qualitative time-interval reasoning, the system enables agents to plan deliberative actions around predictable natural events. The approach is validated through the "Waterway Scenario," demonstrating efficient plan generation in a continuous, concurrent domain.
Introduction
A significant challenge in robotics is the "dynamic world" problem. Classical AI planning typically assumes a "stop-and-go" model where the world remains static between an agent's discrete actions. However, real-world change is often continuous and concurrent. For instance, a barge navigating a canal must account for the rising tide while steering. These "natural actions"—events triggered by physical laws—must be integrated into the agent's reasoning. Previous high-level languages often struggled to model these complexities without succumbing to combinatorial explosion. This work addresses this gap by extending the FLUX language to handle continuous time and simultaneous actions efficiently.
Methods
The proposed system builds upon FLUX, a high-level language rooted in Fluent Calculus. To bridge discrete logic with continuous physical realities, several techniques are employed:
- Constraint Logic Programming (CLP): Utilizing the CLP(R) library allows the system to handle linear constraints over real numbers. Agents reason about time intervals (e.g., an action occurring "at least one hour after state X") rather than fixed points.
- Process Fluents: These state variables represent ongoing processes (e.g.,
Movement) that remain logically stable while representing values that change continuously over time. - Recursive State Update Axioms: To manage concurrency, simultaneous actions are treated as a list, with effects calculated recursively to account for synergic results.
- Qualitative Time Intervals: The planner reduces the search space by focusing on critical intervals: before, during, or after a natural action.
Case Study: The Waterway Scenario
To evaluate the system, a specialized benchmark called the Waterway Scenario was developed. In this scenario, an agent must steer a barge through a system of canals and locks. Success depends on coordinating movement with the tides, as locks are only passable at high tide.
Results
The extended FLUX system successfully navigated the Waterway Scenario. Key results include:
- Plan Generation: The system produced a 4-step plan to reach the harbor.
- Temporal Precision: Instead of rigid schedules, the system provided valid time intervals for actions (e.g., $TA_1 \in [1.0, 2.0]$ and $TA_2 \in [5.0, 6.0]$).
- Optimization: By applying a
Minimize(t)predicate, the system identified the most time-efficient plan within the calculated constraints (e.g., setting $TA_1 = 1.0$).
Discussion and Future Work
While the system effectively integrates natural and deliberative actions, it currently assumes perfect knowledge of when natural events occur. Real-world applications involve uncertainty and sensor noise. Future work aims to incorporate conditional planning, allowing agents to create "if-then" branches based on the actual timing of environmental events, further enhancing robustness in unpredictable domains.
Summary updated