Theseus Maze-Solving Machine
Overview
Theseus was an experimental maze-solving robot I built in 1950 at Bell Labs. It was one of the first examples of a machine that could learn and improve its behavior.
Technical Implementation
The system consisted of:
- The Mouse: A small wheeled vehicle containing magnetic sensors
- The Maze: A 5x5 grid with movable walls and hidden pathways
- Control System: Over 40 telephone relays that stored the learned path
- Power System: Standard electrical power with relay-based logic
How It Learned
- Exploration Phase: Mouse moves randomly, exploring all paths
- Success Detection: When reaching goal, path is electrically recorded
- Memory Storage: Relays store which moves lead to success
- Execution: Subsequent runs use stored path, ignoring dead ends
The relays used were “sticky” - once energized, they stayed that way until manually reset. This created permanent memory of successful paths.
Innovations
- First learning machine: Demonstrated learning without explicit programming
- Relay computation: Pioneering use of telephone equipment for computation
- Artificial intelligence: Early AI before the field had a name
Public Reception
The machine was demonstrated to the press and technical community, generating significant interest. It appeared in:
- Popular Science Monthly
- Scientific American
- Various news outlets
Legacy
Theseus demonstrated that:
- Machines can exhibit intelligent behavior
- Learning can be implemented in hardware
- Intelligence isn’t exclusively biological
The principles influenced later work in AI, robotics, and reinforcement learning.
Theseus proved that with the right circuits, machines could learn.