Introduction: The Tabula Rasa Problem
In traditional robotics, engineers typically provide robots with explicit semantic models of their hardware: "this is a sonar sensor at 30 degrees," or "this motor command translates to forward velocity." Pierce and Kuipers (1994) challenge this paradigm by addressing the tabula rasa learning problem. They ask: can an autonomous agent, starting with no knowledge of its sensors or effectors, learn to navigate and map its environment?
Imagine an agent presented with a control console of unlabelled sliders and a display of flickering bars. Without knowing which bar corresponds to which physical sensor, or how sliders affect movement, the agent must bridge the gap from raw, continuous control signals to a high-level, discrete "cognitive map."
System Architecture
The authors propose a hierarchical framework that abstracts continuous sensorimotor experience into a Finite State Automaton (FSA). This architecture allows the robot to move from low-level control to symbolic reasoning.
The process involves three primary stages of abstraction:
- Feature Learning: Identifying stable, useful signals from raw sensor data.
- Behavior Learning: Discovering degrees of freedom and developing closed-loop control.
- Defining Abstractions: Identifying discrete states and transitions to build a topological map.
Methods: From Raw Data to Features
The first hurdle is making sense of the "sensory input" (often visualized as a raw vector of values). The robot uses a "generate-and-test" approach to find useful features.
Sensor Topology from Correlation
A key insight is that sensors physically close to each other in the environment will show high temporal correlation. By analyzing these correlations and using a relaxation algorithm, the robot can "self-organize" its sensors into a functional spatial arrangement. For example, a robot with 16 distance sensors can discover that they are arranged in a ring without being told their physical mounting positions.
Behavior and Action Models
To understand its motors, the robot employs Principal Component Analysis (PCA) on the effects of random commands. This allows it to identify independent degrees of freedom, such as rotation and translation. It then builds:
- Static Action Models: For open-loop movements toward target features (hill-climbing).
- Dynamic Action Models: For closed-loop error correction, enabling robust path-following (e.g., wall-following).
Experimental Results
The system was validated using a simulated mobile robot in a 10-meter square room. The robot successfully progressed through several "levels of competence":
- Random Wandering: Initial state with no learned models.
- Open-Loop Navigation: Using static models, though prone to drift.
- Closed-Loop Navigation: Using dynamic models to reliably follow the room's perimeter.
Ultimately, the robot successfully abstracted its continuous environment into a discrete set of states—nodes representing path intersections or terminations—which could then be used by high-level AI agents for map building and path planning.
Conclusion and Future Work
This work demonstrates that a robot can autonomously discover the structure of its own body and environment. While the experiments were conducted in relatively simple, static environments, the "ladder of abstractions" provides a robust blueprint for autonomous discovery. Future challenges include scaling these methods to thousands of sensors and more complex, dynamic real-world settings.
Summary updated