Abstract: This paper explores the evolution of AARON, an autonomous drawing system that transitioned from abstract shapes to complex scenes of humans and plants. Developed by Harold Cohen, AARON mimics human cognitive strategies to generate plausible art from sparse, non-visual world knowledge. By utilizing a hierarchical knowledge base and feedback-driven execution, the system achieves autonomous complexity without traditional computer graphics techniques like ray-tracing or perspective projection.
Introduction
The fundamental challenge in autonomous art is determining what a system must "know" about the world to create a plausible visual representation. Traditional computer graphics often rely on "knowledge-free" mechanical processes like ray-tracing, which lack an understanding of the objects being rendered. In contrast, AARON investigates the cognitive principles of "visual imagining," moving toward realistic representations of humans and plants by managing occlusion, anatomy, and spatial planning autonomously.
Methodology
AARON employs a sophisticated blend of rule-based logic and feedback-driven execution. Its knowledge is organized into a five-level hierarchy:
- Declarative: Basic anatomy (e.g., a human has a head, torso, and limbs).
- Exemplary: Prototypical dimensions (e.g., "long" arms vs. "large" heads).
- Structural: Where the joints/articulations are located.
- Functional: The legal range of motion for those joints.
- Behavioral: How the figure maintains balance and performs gestures.
A key strategy is the "Conceptual Core." Similar to child development, AARON first generates a "stick figure" or "scribble" (the core) and then "embodies" it by running a maze-running procedure to draw a path around the marked cells of the core.
This process is feedback-dependent; the system does not pre-plan lines but adjusts its sampling rate based on the importance of the part being drawn—drawing a hand with high precision and a thigh more loosely. To manage spatial complexity, AARON uses a hybrid 2D/3D planning approach, maintaining a 3D "floor plan" to ensure physical plausibility before projecting the scene back to the 2D drawing plane.
Results and Insights
By 1987, AARON was capable of generating unique, non-repetitive drawings of multiple people in dense garden settings without human intervention. This shift from simple line generation to object embodiment allowed for complex scenes involving occlusion and depth.
A central insight of the project is that "plausibility" is more important than mathematical "correctness." A representation is valid if it is cognitively recognizable, even if anatomically inaccurate. Furthermore, Cohen argues that "style" is an emergent property—the "signature of a complex system"—arising from the interaction of thousands of interdependent decisions within the program's architecture.
Discussion and Limitations
Despite its successes, AARON faces technical limitations. It is currently optimized for "blob-like" forms or those with a central axis, struggling with "cube-like" geometric objects. Its knowledge is also domain-specific, requiring manual coding for new object types (e.g., flora, humans, geology). Future work aims to develop a more dynamic model of "imaginational planning" to replace the current static 3D floor plan, allowing for even more sophisticated spatial reasoning.
Summary updated