Ian Horswill's 1993 work at the MIT AI Lab introduced Polly, a minimalist robot designed to navigate office environments. At a time when robotics research focused on complex 3D reconstructions, Polly demonstrated "machinery parsimony"—the idea that a robot can navigate robustly by exploiting the specific structural constraints of its environment to simplify computation.
The Problem: Computational Overload in 90s Vision
In the early 1990s, computer vision for robotics was typically "bottom-up," attempting to build complete 3D maps of the world before making navigation decisions. This approach was computationally expensive, slow, and often failed in messy, real-world settings. Horswill addressed the challenge of efficient, task-based navigation by asking: What is the minimum amount of information a robot needs to perform its job? Instead of solving general vision, Polly solved the specific problem of office navigation by leaning on predictable building geometry.
Methodology and Architecture
Polly’s design is based on Active Vision and Environmental Constraints. The system eschews laser rangefinders and stereo cameras in favor of a single low-resolution (64x48) grayscale camera.
Key technical approaches include:
- Step-wise Transformation: A methodology where complex algorithms are simplified into specialized ones by applying environmental assumptions.
- Radial Depth Mapping (RDM): Instead of calculating true 3D coordinates, Polly uses the "ground-plane constraint" to treat image height as a proxy for distance.
- Qualitative Navigation: Polly uses a topological map rather than precise coordinates. It recognizes landmarks (like the kitchen or elevator lobby) and uses simple "difference reduction" to move between them.
- Parallel Architecture: The software consists of independent Scheme procedures running in a loop, mimicking a hardware circuit to ensure real-time responsiveness.
Environment and Evaluation
Polly was not trained on modern datasets like ImageNet; instead, its "dataset" was the 7th floor of the MIT AI Lab. The robot was evaluated through hundreds of hours of service in this live, unmodified environment.
Key Insights: The Environment as a "Compiler"
Horswill argued that the physical structure of the world—such as flat floors and straight hallways—acts like a compiler that optimizes a robot's code.
- The Ground-Plane Constraint: On a flat floor, the further away an object is, the higher up it appears in the camera's field of view. This allows for obstacle avoidance via simple image processing.
- The Background-Texture Constraint: In the MIT lab, the carpets were untextured. Therefore, any detected "edges" were assumed to be obstacles, turning a difficult recognition problem into a simple edge-detection task.
- Negative Information: Polly’s "unwedger" unit assumed the robot was stuck if it failed to move forward for 2 seconds, triggering an automatic turn.
Results and Performance
Polly achieved remarkable efficiency for its era:
- Speed: The system ran at 15Hz, updating every 66ms.
- Efficiency: The entire navigation system was written in fewer than 1,000 lines of Scheme code and ran on a TI C30 DSP with only 64KW of RAM.
- Robustness: Polly could navigate at speeds up to 1.5 meters per second and recognized the lab's kitchen with 80-100% accuracy.
Limitations
Despite its success, Polly's minimalist approach had trade-offs. Place recognition was its weakest link; moving a chair or a person standing in front of a landmark could cause the robot to get lost. Additionally, the system was prone to "hallucinations," occasionally braking for shadows or reflections on polished floors. Because the high-level sequencers were rigid scripts, the robot lacked the ability to re-plan if it missed a turn, relying instead on wandering until it recognized a new landmark.
Summary updated