Introduction
Traditional AI planning systems, exemplified by frameworks like STRIPS, historically operated under "closed world" assumptions. In these environments, the state of the world remains static unless altered by the agent's direct actions. However, real-world domains—most notably autonomous driving—are inherently dynamic. In such settings, environmental conditions change independently of the agent, and other actors operate with their own intentions. This paper introduces AUTODRIVE, a simulation framework designed to bridge the gap between high-level planning and the unpredictable physics of a multi-agent traffic environment.
The Challenge of Dynamic Environments
The core difficulty in dynamic planning is that an agent lacks perfect information and total control. A planner cannot simply update its internal world model as a matter of fact; it must instead perceive the world through snapshots and deduce the state of its surroundings. AUTODRIVE addresses this by decoupling the Planner from the World Simulator. The simulator acts as a mediator, enforcing physical constraints—such as momentum, braking distances, and road friction—that the planner may not explicitly model but must nonetheless navigate.
Methodology: The AUTODRIVE Framework
Implemented in PROLOG, the system utilizes several key mechanisms to manage the complexity of a dynamic world:
Spatial Segmentation and Segment Controllers
To maintain computational efficiency, the simulated environment is divided into discrete "segments," such as specific road stretches or junctions. Each segment is managed by a Segment Controller. This controller acts as an intermediary for perception. Instead of providing the planner with a global "god-eye view," the controller determines which objects are within the agent's line of sight. Objects within a segment "shout out" their presence to the driver, allowing for realistic occlusion. For instance, a large vehicle like a bus can physically block the "shouts" of objects further down the road, naturally simulating visibility limitations.
Action Mediation and Physics Enforcement
In AUTODRIVE, the planner does not move itself. Instead, it sends an "intended action" to the simulator. The simulator evaluates this intention against the current physical state (e.g., velocity, presence of hazards like oil spills). The simulator then returns the actual resulting position to the planner. This feedback loop forces the planner to reconcile its expectations with physical reality, learning to handle environmental "side-effects."
Temporal Unitization and Intent Deduction
Time is divided into discrete units. By comparing environmental snapshots across these units, the planner can calculate the speed and direction of other agents. This allows for a rudimentary form of "theory of mind," where the planner hypothesizes the intentions of other drivers to predict and avoid potential collisions.
Experimental Environment
The framework was tested using a digital reconstruction of real-world locations in Brighton, UK. The simulation map includes:
- Lewes Road: The primary corridor for testing.
- Key Junctions: Inverness Road, Aberdeen Road, and Edinburgh Road.
- Regulatory Elements: The database incorporates specific road markings, including double yellow lines, broken white lines, and "give-way" signs, ensuring the agents must operate within a rule-bound framework.
Key Insights and Results
The PROLOG implementation successfully demonstrated several critical capabilities:
- Multi-Agent Interaction: Multiple AUTODRIVE instances can operate simultaneously, creating a complex environment of interacting "thinking" drivers.
- Rule-Bound Navigation: Agents successfully navigated junctions and performed maneuvers like overtaking while adhering to established traffic laws and rights of way.
- Perceptual Realism: By providing data relative to the "observer" (the car's front bumper) rather than absolute coordinates, the system accurately modeled the challenges of ego-centric navigation.
Limitations and Future Directions
While the segmentation model provides efficiency, the "shouting out" mechanism may face scalability issues in high-density environments. Furthermore, the current model assumes a level of uniformity in driving behavior. The author suggests that future work should incorporate driving "personalities" to reflect diverse styles and environmental stressors like fog, which would require the segment controller to dynamically veto information based on visibility thresholds.
Summary updated