The Relay Maze Mouse
How Theseus Remembered Fifty Relay Bits of Trial and Error
A hollow wooden mouse carrying a small two-inch bar magnet inside its body sat at the corner of a twenty-five-square aluminum labyrinth at Bell Telephone Laboratories in 1950. Named Theseus after the hero of Greek mythology who navigated the Cretan labyrinth of King Minos, the life-sized rodent had copper wire whiskers protruding from its nose. It rolled smoothly across the floor on three small brass wheels. Yet inside the carved wooden shell there were no batteries, no motors, and no electronics whatsoever. All of the sensory routing, spatial memory, and motor actuation lived hidden beneath the maze floor inside a cabinet of clicking telephone relays.
Between 1950 and 1952, Claude Shannon built Theseus in his workshop to demonstrate how a small switching network could exhibit trial-and-error exploration, spatial memory, and loop-breaking adaptation. The top surface of the apparatus consisted of a five-by-five checkerboard of twenty-five square cells. Thin aluminum partition walls could be plugged into slots between any adjacent squares to create a new maze pattern in seconds. A brass terminal piece shaped like a wedge of cheese could be placed in any target square.
Directly underneath the thin non-magnetic floor rode a motorized carriage driven along perpendicular $X$ and $Y$ lead screws like an upside-down crane. Mounted on that carriage was an electromagnet facing upward. When the electromagnet moved and rotated beneath the floor, magnetic flux pulled the bar magnet inside Theseus along in lockstep above the surface. Whenever the mouse's copper whiskers brushed against an aluminum partition wall, the grounded metal closed a low-voltage sensing circuit. That pulse signaled the relay brain below that the attempted corridor was blocked.
As Shannon explained at the eighth Macy Cybernetics Conference in New York in March 1951, the entire learning mechanism relied on approximately seventy-five to ninety standard Bell System telephone relays. Those switches were divided into two groups: a small control unit and a fifty-bit spatial memory array. Because the maze had twenty-five squares and four cardinal compass directions—North, South, East, and West—Shannon wired exactly two latching memory relays to each square. Two binary relays can rest in $2^2 = 4$ states (00, 01, 10, 11). That two-bit register allowed each square to remember the single compass direction in which the mouse most recently exited that cell.
When dropped into an unfamiliar maze, Theseus executed a systematic search strategy. Upon entering a square, the relay control read that cell's two memory bits, rotated the mouse ninety degrees counterclockwise from the stored direction, and advanced. If the copper whiskers hit an aluminum wall, the mouse backed up to the center of the square, rotated another ninety degrees, and tried the next direction. The moment the mouse succeeded in leaving a square through an open doorway, the two relays assigned to that square latched to record the successful exit vector.
That two-bit-per-square overwrite rule solved two classic artificial intelligence problems at once. First, once Theseus touched the brass goal terminal, the chain of stored exit arrows pointing from the starting square to the cheese formed a loop-free shortest path. Any dead-end detours explored along the way had been automatically overwritten the last time the mouse stepped out of the junction square. Second, if an experimenter moved an aluminum wall to trap Theseus in a closed loop, the ninety-degree rotation rule forced the mouse to break out of the cycle rather than circling forever. When Bell Telephone Laboratories filmed Theseus in 1952 for its archival documentary on machine learning, the wooden mouse solved a newly rearranged twenty-five-square labyrinth by trial and error and then sprinted from start to goal on its second run in twelve seconds without a single wrong turn.