DARPA 自动驾驶挑战赛

2004 年 3 月，美国莫哈维沙漠。DARPA 举办的自动驾驶挑战赛鸣枪，15 支车队把各自的无人车送上 142 英里赛道，结果令人尴尬——没有一辆车完成全程，最远的也只跑了大约 7 英里。沙尘、侧风、崎岖地形，把这些实验室里精心调校的车辆打得七零八落。但这场「失败」恰恰是自动驾驶产业的真正起点。

DARPA 办这个比赛的初衷很直接：用公开竞赛代替闭门研发，逼出可用的无人车技术。2001 年立项，2004 年首赛，规则简单粗暴——谁能最先让车辆自动跑完全程谁赢。没有华丽的演示，没有安全员坐在方向盘后，只有沙漠里的真实路况。这种「用真实世界当考场」的思路，是自动驾驶第一次以工程命题而非学术演示的形式出现。

转机发生在第二年。2005 年的第二届比赛，5 辆车成功跑完全程，冠军斯坦福的 Stanley 用激光雷达加机器学习，在 6 小时 53 分钟内完成了 132 英里。从「零完赛」到「五车完赛」，只隔了一年。这一年的差距不是运气，而是感知、规划、控制整个技术栈被沙漠逼出来的快速迭代。

2007 年第三届比赛换了个考场：一座废弃空军基地改造的模拟城市。车辆要在街区里行驶、避让其他车辆、遵守交通规则、处理交叉路口——在动态交通里连续运行 60 英里。这一届比的不再是「能不能自动驾驶」，而是「能不能在有人类社会的规则下自动驾驶」。许多后来著名的自动驾驶团队，都是从这三届比赛里走出来或受其启发的。

回看这三年，DARPA 挑战赛最珍贵的地方，是把「自动驾驶」从一个遥远设想变成了一个有明确规则、有真实环境、有公开成绩的工程赛道。今天的 Waymo、特斯拉 FSD，乃至整个自动驾驶产业，都能在这三场沙漠与城市竞赛里找到自己的最早源头。失败的首届并不丢人——它证明这条路需要重来，而重来的方式，是把技术放到真实的路上检验。

In March 2004, in California's Mojave Desert, DARPA held its first Grand Challenge. Fifteen teams sent their unmanned vehicles onto a 142-mile course. The result was embarrassing: not a single car finished, and the best run covered about seven miles. Dust, crosswinds, and rough terrain shredded carefully tuned laboratory vehicles. Yet this "failure" is where the self-driving industry truly begins.

DARPA's motive for the race was blunt: use open competition instead of closed-door research to force usable autonomous-vehicle technology. The program began in 2001, the first race ran in 2004, and the rules were simple—whoever's vehicle could cover the course on its own won. No polished demos, no safety drivers behind the wheel, just real desert conditions. This "the real world is the exam" approach was the first time self-driving appeared as an engineering problem rather than an academic demonstration.

The turn came a year later. In the 2005 race, five vehicles finished; Stanford's Stanley, using lidar and machine learning, covered 132 miles in about 6 hours 53 minutes. From zero finishers to five in a single year was not luck—it was the entire perception-planning-control stack being forced into fast iteration by the desert.

The third race, in 2007, changed the venue: a mock city built on a decommissioned air base. Vehicles had to drive through blocks, avoid other traffic, obey traffic rules, and handle intersections—about 60 miles of continuous operation amid dynamic traffic. This round was no longer about "can you drive autonomously" but "can you drive autonomously under society's rules." Many later-famous self-driving teams emerged from or were inspired by these three races.

Looking back, the most precious thing DARPA's challenges gave us was turning "self-driving" from a distant idea into an engineering field with explicit rules, real environments, and public results. Today's Waymo, Tesla FSD, and the entire autonomous-driving industry can trace their earliest roots to those three races in the desert and the city. The failed first race was not shameful—it proved the road needed to be retried, and the way to retry it was to test technology on real roads.