In March 2004 DARPA offered a 1 million dollar cash prize to any team whose vehicle could drive itself across roughly 142 miles of the Mojave Desert with no human at the wheel. Fifteen vehicles started. None finished. In DARPA’s own later account, “the top-scoring vehicle traveled only 7.5 miles” before getting stuck, and “the prize went unclaimed.” Carnegie Mellon’s Red Team, the frontrunner, hung up on a rock after a switchback turn.
By every conventional measure it was a flop. A million-dollar prize, a year of work by serious teams, and the best car covered about five percent of the course. Press coverage at the time was unkind.
Then the remarkable part happened. DARPA reran the challenge in 2005, and this time five vehicles out of the field completed a 132-mile course in southern Nevada, with Stanford’s “Stanley,” led by Sebastian Thrun, winning the now-doubled 2 million dollar prize. A complete failure to a clean sweep in eighteen months is one of the fastest failure-to-success turnarounds in the history of AI.
DARPA’s retrospective argues the 2004 event was not wasted: it “offered a promising glimpse at what was possible” and “created a community of innovators” that drove “major advances in the development of autonomous robotic ground vehicle technology.” The teams and engineers forged in those desert races went on to staff Google’s self-driving project and much of the autonomous-vehicle industry that followed. The lesson is that in fast-moving fields, a public, well-defined failure can be worth more than a quiet success - it shows exactly where the hard problems are and rallies people to solve them.