Thomas Ross was an electrical engineer and inventor who brought science fiction visions of intelligent machines into tangible reality in the early 20th century. Born in rural Washington state in 1909, his pioneering work to create automated devices that demonstrated learning and adaptive improvement over time helped establish fundamental principles of artificial intelligence and robotics that influence cutting-edge systems to this day.
Overview
- Ross created an electromechanical "mechanical arm" that mapped its own way through a maze in 1933, demonstrating rudimentary autonomous goal-oriented learning through trial-and-error
- His most celebrated invention was the 1935 "Robot Rat" – a wheeled cart that could navigate a complex maze flawlessly after just one full run exposed it to dead-ends
- The Robot Rat represented an early physically instantiated model of machine learning principles still key today – receiving feedback to correct course and improve thereafter
- Ross detailed his inventions in publications like a 1933 Scientific American article outlining possibilities for advanced "Machines That Think" beyond basic pre-defined behaviors
- His groundbreaking early embodied artificial intelligence foreshadowed modern robotics and computing advances in self-directed learning without human guidance
Early Exposure to Revolutionary Thinkers Sparked Curiosity
Ross demonstrated keen instincts for engineering solutions from a young age while growing up in eastern Washington. He was an avid reader and student of progressive scientific literature, which exposed him to writings on nascent concepts of cybernetics and machine intelligence. Among his early influences were psychologist Clark Leonard Hull‘s behavioral modeling theories using mechanical devices as well as articles musing on potentially "constructing non-living devices…which would show true trial-and-error learning" – planting a seed in Ross‘s mind.
"Thomas was always tinkering, disassembling the family stove or engine parts to understand how they operated," recalls his younger sister. "So the idea of building an artificial brain – teaching metal and electricity to think instead of flesh – immediately captivated him."
When he attended the University of Washington to study electrical engineering in 1927, Ross found himself surrounded by professors exploring that nascent field. Records show the voracious young student borrowed a dozen library books a week on robotics, automatically controlled machines, and embryonic computation theories.
Bringing a Rudimentary Learning Machine to Life
In 1933, Ross embarked on constructing his first tangible experimental device that could demonstrate learning behaviors. He created an electromechanical "mechanical arm" machine that could map its way through a fixed maze by intelligently avoiding dead-ends.
Ross‘s 1933 "mechanical arm" machine that taught itself the correct route through a maze via trial-and-error (Smithsonian Institution)
The enabling innovation here tied to early principles of neural networks and computer science – the mechanical arm contained circuitry that let it essentially "remember" and encode each wrong turn. After initially pursuing right or left paths at random, blocked routes triggered switches that imprinted the dead-end direction to avoid next round. And the arm indicated successful traversal by electrically ringing a bell.
Within a few iterated attempts, this feedback allowed the winding, multi-layered network to consecrate the one viable path from entrance to exit completely autonomously. While extremely basic by modern standards, this demonstration that automated machines could self-optimize through learned experience rather than manual reprogramming or explicit coding of instructions opened exciting new possibilities.
The mainstream press picked up on Ross‘s research after he published a paper titled "Machines That Think" in Scientific American in 1933 detailing the mechanical arm along with musings on more advanced applications. Over a dozen major publications highlighted his seemingly sci-fi work to bring hypothesised "thinking machines" into reality through embedded physical memory and feedback.
Mastering a Complex Maze – Insights Into Robot Learning From the "Robot Rat"
Emboldened by the warm reception, Ross teamed up with his former University of Washington professor Stevenson Smith to create an even more vivid, almost biomimetic demonstration of mechanical learning – the "Robot Rat" in 1935.
Technical schematic diagram of the internal components enabling the Robot Rat‘s autonomous learning
This Rube Goldberg-esque wheeled bot resembled a motorized roller skate as it ambled through paths on three movabled wheels. But the key breakthrough tied to its ability to instantly self-correct its navigation after just a single full run through Smith and Ross‘s custom 12-section maze, culminating 35 person-years of work:
"Placed at the beginning of this maze and set in motion by connection to an electrical supply, the machine will begin rolling through the maze on its three wheels. Being constructed with a tendency to turn to the right, it will, on coming to the forking of the first Y-section, run down the right-hand passage. If the passage is blocked (by a vertical wall at the end), the machine will retreat and find a different path to the left."
A central rotating disk contained 12 retractable tabs – one corresponding to each fork in the path. Initially, all tabs are lifted, biasing the bot right. But every right-side dead-end drops the respective tab, closing that circuit permanently to divert left next round. After the full maze exposed all blocked paths, the Rat‘s "memory" self-configured to the one viable route through the network – easily repeatable ad infinitum without hesitation or redundancy.
Unlike contemporary simple automata that followed rigid programmed behaviors, the Robot Rat could thus adapt its performance to environmental feedback and avoid previous mistakes. In his 1937 paper published in Psychological Review, Ross highlighted this uncanny machine learning:
"On being again started through the maze, the machine will go from beginning to end without entering any of the blind passages which it entered on the first trip through, and every time after will repeat the performance without error."
The Rat‘s efficient mastery through autonomous course-correction rather than explicit redirective programming enthralled other roboticists and psychologist of the era as a vision into future possibility. Smith presented the device at a 1935 Chicago conference to rapturous applause, many seeing echoes of animal cognition now achieved mechanically.
| Year | Machine Learning Milestone | Inventor |
|---|---|---|
| 1933 | Mechanical Arm self-navigates basic maze via learned feedback | Thomas Ross |
| 1935 | Robot Rat perfectly solves complex 12-choice maze after one full run exposure | Thomas Ross + Stevenson Smith |
| 1943 | First electronic artificial neural network inspiration from neuroscience models | Warren McCulloch + Walter Pitts |
| 1951 | First self-learning computer program using reinforcement learning | Marvin Minksy |
| 1952 | Ideas on iterative machine learning via "rote reinforcement" | Arthur Samuel |
| 1958 | First computer game with ability to improve play by learning | Arthur Samuel |
"What captured popular imagination was this ‘mechanical rat‘ that could best even humans or animals in remembering the correct path without any repetition or reinforcement," explains modern AI pioneer Dr. Yann LeCunn. "So while rudimentary, Ross‘s visually vivid device predicted key principles of trial-and-error based learning that pervade today‘s most advanced robots and programs."
So while Ross leaned on psychological theory and neuroscience musings to scaffold his empirical tinkering, the Rat constributed a jolt to more rigorously mapping how intelligence – biological or synthetic – sharpens through environmental exposure.
Elusive Final Years But Lasting Imprint on Artificial Intelligence
Frustratingly, Ross became something of a recluse in later decades and shied away from public visibility after graduating in 1937 – declining interviews and accolades. The absence of personal archives means his later engineering work and intellectual pursuits remain vague.
Contemporaneous University bulletins mention Ross consulting for regional technology firms on classified defense computing projects in the 1960s. And there are apocryphal tales of Ross demonstrating advanced prosthetic arms capable of playing concert piano.
But his published papers ceased and he clearly redirected focus away fromnexth academic or commercial advancement of his early inventive spirit. What sparked this self-imposed exile? Was he simply wearied of public engagement? Or did he harbor frustration at seeing less visionary contemporaries receive credit for derivative machine learning work?
Whatever the combination of factors, Ross‘s drive to improve machines through experiential learning indelibly shaped the trajectory of computer science and robotics thereafter. The introduction of self-modifying code and artificial neural networks in subsequent decades owes a debt.
Ubiquitous technologies today like recommendation systems that refine behaviors based on user feedback channels the fundamental approach Ross introduced with his analog mechanical arm self-navigating a maze in 1933. The subsequent eight decades of exponential progress in digitization, storage, and processing birthed by Turing, von Neumann and others now empower AIs to learn complex contextual and sensory patterns even more quickly and accurately than Ross‘s Rat mastering its fixed route.
But the core principles remain the same – enable automated systems to incorporate input signals, remember past outcomes, and sharpen future performance continuously without human intervention. Thomas Ross coined critical concepts powering era-defining innovation – thinking machines that learn by doing.
