Introduction
In the early 1980s, AI planning systems often operated under a "closed-world assumption," where the environment was static and actions were guaranteed to succeed. However, real-world navigation—exemplified by taxi driving—is inherently dynamic. Unexpected obstacles, missed signs, and changing conditions require a system that can monitor its own execution and recover from failures. This paper presents ELMER, an agent designed to navigate a simulated city by integrating hierarchical planning with robust error detection and recovery strategies.
The ELMER Architecture
The ELMER system is composed of three interdependent modules: the Planner, the Map, and the Executor. Unlike systems that treat these as separate stages, ELMER treats planning, execution, and learning as a continuous loop.
- The Planner: Responsible for generating routes by "splicing" together known path segments from the Map to meet specific customer goals.
- The Map: A hierarchical knowledge base containing routes, landmarks (e.g., buildings, stop signs), and "associations" that describe the spatial relationships between different areas.
- The Executor: The runtime component that carries out the plan. It matches low-level sensory "windows"—what the agent currently sees—against the expectations defined in the active plan.
Strategies for Error Detection
ELMER does not assume a plan will work; it actively looks for evidence of failure. The system employs three primary detection methods:
- Explicit Error Transitions: The planner can embed conditional checks directly into the plan hierarchy. If a specific landmark is observed that should not be visible on the current path, an error is triggered.
- Sensory Monitoring: The Executor compares real-time sensory data with historical data from previous successful traversals. Discrepancies in permanent features (like street signs) signal that the agent has veered off-course.
- Distance Tracking: Each plan segment includes an expected travel distance. If the agent exceeds this threshold without reaching the next transition point, it concludes it is lost.
Error Recovery and Learning
Once an error is detected, ELMER attempts to recover using a tiered approach. It may try retracing its steps by reversing recent actions, though this is limited by constraints like one-way streets. A more robust method is hierarchical localization: the agent "zooms out" to a higher level of the plan hierarchy to find a recognizable regional landmark or a nearby route associated with its general vicinity. If these fail, ELMER resorts to exploration, moving in the general direction of the goal until it encounters a known territory.
Crucially, ELMER views failure as a learning opportunity. When a plan fails, the erroneous path is abstracted and added to the Map as an "error transition." This allows the system to recognize and avoid similar mistakes in future planning cycles.
Evaluation and Insights
The system was tested in "Simon City," a simulated microworld featuring both permanent landmarks and transient obstacles like moving cars. While the paper focuses on architectural proof-of-concept rather than quantitative metrics, it demonstrates that hierarchical plans serve as a vital safety net. By distinguishing between transient features (pedestrians) and permanent ones (buildings), ELMER maintains a stable sense of location even in a noisy environment.
Future work aims to address limitations such as the lack of a global coordinate system and the computational cost of matching sensory windows against a growing map. Nonetheless, ELMER provides a foundational model for autonomous agents that must navigate the complexities of the real world.
Summary updated