The Grand Tour: A Casual History of Autonomous Vehicles

Imagine a world where cars drive themselves. Not just on test tracks, but on the chaotic, unpredictable streets we navigate every day. This isn't a distant sci-fi dream; it's the reality being engineered today, with robotaxis ferrying passengers in San Francisco and driver-assist systems navigating highways worldwide. But to understand the full-stack symphony of hardware, software, and data that makes this possible—and why it’s built the way it is—we must first rewind the tape. Way back. To an era before GPUs, before lidar, before the term “AI” meant anything more than a academic curiosity. This chapter is your grand tour, a casual yet deeply researched history that connects the dots from radio-controlled phantoms to the autonomous vehicle (AV) stacks of today. We’ll spotlight the inflection points, celebrate the unsung heroes, and dissect the spectacular failures that taught us everything. Because the story of autonomous driving isn’t just a timeline; it’s a masterclass in iterative failure, visionary persistence, and the relentless pursuit of turning “impossible” into “shipped.”

The Pre-Computer Age: Phantoms, Predictions, and Pipe Dreams (1920s–1970s)

The desire to decouple driving from human intervention is almost as old as the automobile itself. The first documented “autonomous” vehicle wasn’t a computer on wheels, but a clever hack. In the 1920s, engineers at the Radio Corporation of America (RCA) demonstrated a “phantom auto” in a controlled environment. A driver, hidden in a following car or a control booth, used a radio antenna mounted on a standard 1920s sedan to steer, accelerate, and brake remotely. The car, guided by pulses sent to a receiver on the dashboard, could follow a predetermined path. It was a parlor trick, not autonomy, but it captured the public imagination and planted a seed: perhaps the driver could someday be removed entirely. These demonstrations, covered in *Popular Science* and *The New York Times*, framed the conversation for decades—autonomy as remote control, not independent intelligence.

The theoretical groundwork was being laid in parallel. In 1948, science fiction author and futurist Arthur C. Clarke published “The Possibility of Autonomous Ground Transportation” in the *Journal of the British Interplanetary Society*. He speculated about vehicles using “radio beams” or “magnetic induction” to follow paths, concepts that would later resurface as dedicated short-range communications (DSRC) and magnetic guidance for trains. More concretely, in 1956, General Motors unveiled the Firebird II concept car at the Motorama show. Its most stunning feature? A “electronic guide system” that could, in theory, read special cables embedded in the road and control steering and speed. GM’s vision was a controlled, infrastructure-dependent future—a “safety highway” where cars moved in platoons under electronic supervision. It was a beautiful, utopian vision, but it relied on a complete rebuild of the nation’s roadways, a political and economic non-starter. The lesson was early and harsh: autonomy that depends on massive, centralized infrastructure changes is dead on arrival. The winning path would have to be self-contained.

The 1960s saw the first attempts at *onboard* sensing. In 1961, researchers at the University of Illinois presented a prototype that used buried wires and magnetic sensors to guide a modified electric car on a fixed path. It was essentially a robotic train for roads. More ambitious was the work of James Adams at Stanford Research Institute (SRI). In 1969, he developed “Stanford Cart,” a small, six-wheeled rover that could follow a white line on a floor using a simple camera and analog circuitry. It was painfully slow (moving one meter every 10–15 minutes), but it proved a camera could be used for closed-loop control. The cart’s greatest contribution wasn’t its speed but its proof of concept: vision for navigation was possible, if not practical. The era was defined by analog electronics, simple sensors, and a fundamental belief that the problem was primarily one of guidance along pre-defined paths. The complexity of open-world, unstructured environments—pedestrians, weather, ambiguous signage—was barely acknowledged. The dream was elegant; the reality, not yet confronted.

The Dawn of Digital Autonomy: Computers Enter the Fray (1970s–1980s)

The 1970s marked the transition from analog gadgets to digital computation, and with it, a shift from “guided vehicles” to “autonomous vehicles.” The defining project of this era was the Stanford Research Institute’s (SRI) “Cart”, now upgraded with a more powerful computer. In 1979, under the leadership of Hans Moravec, the cart successfully navigated a room filled with obstacles using stereo vision. Moravec’s 1980 PhD thesis, *“Obstacle Avoidance and Navigation in the Real World by a Seeing Robot Rover,”* became a foundational text. He introduced the concept of “sparse depth maps” from stereo vision and a hierarchical control system that separated low-level reflexive obstacle avoidance from high-level path planning. This “sense-plan-act” loop, with its clear modularity, became the architectural DNA for decades of robotics, including AVs. Moravec’s work was pure, academic robotics—slow, fragile, but intellectually rigorous.

Across the Atlantic, a parallel, more audacious effort was underway. Ernst Dickmanns, a professor at the University of Munich, and his team built the VaMP (Versuchsfahrzeug für autonome Mobilität und Prävention) and later the Parma vehicle. Starting in the mid-1980s, they equipped a Mercedes 500 SEL with an array of sensors—cameras, lidar-like laser range finders, and inertial systems—and, crucially, a spatiotemporal world model. Dickmanns’ team pioneered “behavior-based” control with a focus on “dynamic vision”: processing only the relevant parts of the scene at any given moment (e.g., the road ahead, distant traffic) to manage computational limits. In 1994 and 1995, their vehicles achieved stunning feats: driving over 1,000 kilometers on French autoroutes at speeds up to 130 km/h, with minimal human intervention. The 1995 run, from Munich to Copenhagen and back, included lane changes, following curves, and handling oncoming traffic.

Dickmanns’ contribution was monumental and, for too long, underappreciated in the Anglo-American AV narrative. He demonstrated that high-speed autonomy on public roads was possible with the technology of the era, provided you had exquisite sensor fusion and a robust world model. His approach was not about raw perception but about attentive, anticipatory control. The VaMP’s system could track multiple moving objects, predict their trajectories, and plan maneuvers accordingly—a direct precursor to modern motion forecasting modules. Yet, his work was largely confined to European research circles. The “unsung hero” narrative is strong here: Dickmanns solved the core problem of real-time, high-speed perception and control years before the DARPA challenges made headlines. The reason his work didn’t immediately spark an industry was twofold: the computational hardware was still prohibitively expensive and custom-built, and there was no clear commercial pathway or massive funding push. It was a brilliant academic proof, waiting for its moment.

Meanwhile, in the United States, the NavLab project at Carnegie Mellon University (CMU), led by Takeo Kanade and Chuck Thorpe, was pursuing a different philosophy. Starting in 1984, NavLab vehicles like the “Navlab 1” (a modified Chevy van) and later the “No Hands Across America” (1995) journey used a “sensor fusion” approach, combining cameras, laser scanners, and inertial navigation. The 1995 cross-country trip from Pittsburgh to San Diego—a 2,850-mile journey—was a landmark. The vehicle, Navlab 5, drove 98.2% of the distance under computer control, with a human ready to take over. Average speed was a glacial