# How did Stanford's robot car Stanley win the 2005 DARPA Grand Challenge?

Date: 2020-11-18 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/how-did-stanley-win-the-darpa-grand-challenge/

**Answer:** On 8 October 2005, Stanley, a driverless Volkswagen Touareg built by a Stanford-led team, drove itself across about 132 miles of Mojave Desert in just under seven hours and won DARPA's $2 million Grand Challenge. It won largely because it learned from data what safe, drivable ground looks like, instead of relying only on hand-written rules.

Nineteen months after every robot in DARPA's first Grand Challenge broke down within a few miles, the agency held the race again. This time the result was completely different. On 8 October 2005, five driverless vehicles crossed the finish line of a desert course of about 132 miles (212 km), and the first to arrive was Stanley, a Volkswagen SUV covered in sensors, entered by Stanford University.

Stanley's winning time was 6 hours 53 minutes and 58 seconds according to the team's own published paper; Wikipedia lists 6:53:08. Either way, it averaged 19.1 mph over rough ground with no human help, and earned its team the $2 million prize.

## The second race

DARPA, the research agency of the US Department of Defense, doubled the prize after the 2004 race went unclaimed. Interest grew too. The Stanford team's account reports that 195 teams registered. After site visits, 43 semifinalists were invited to a qualifying event at the California Speedway in Fontana, California, held from 27 September to 5 October 2005. Stanley was the only vehicle to pass all 50 gates and avoid every obstacle on all four of its qualifying runs, which earned it the second starting position.

Twenty-three vehicles started the race. The course both began and ended in Primm, Nevada, and wound through three tunnels, more than a hundred sharp turns and, at the end, Beer Bottle Pass, a narrow mountain road with steep drops on both sides. As before, teams received the route only two hours before the start, as a file of 2,935 GPS waypoints with a speed limit for each section.

Stanley started at 6:35 a.m. (Wikipedia gives 6:40). DARPA staff paused it twice so it would not run into Carnegie Mellon's robot H1ghlander, which had started five minutes earlier. At mile 101.5, officials paused H1ghlander instead and let Stanley pass. The final order:

| Place | Vehicle | Team | Time (Wikipedia) |
|---|---|---|---|
| 1 | Stanley | Stanford Racing Team | 6:53:08 |
| 2 | Sandstorm | Carnegie Mellon University | 7:04:50 |
| 3 | H1ghlander | Carnegie Mellon University | 7:14:00 |
| 4 | Kat-5 | Gray Team | 7:30:16 |
| 5 | TerraMax | Oshkosh Truck | 12:51, over the ten-hour limit |

The rivalry had a personal edge. Sebastian Thrun, who led the Stanford team, had previously been a professor at Carnegie Mellon and a colleague of Red Whittaker, who led the Carnegie Mellon effort.

## What was inside Stanley

Stanley was a 2004 Volkswagen Touareg R5 TDI, a diesel SUV, adapted with help from Volkswagen of America's research lab. On its roof sat five laser range finders, each sweeping a thin beam across the ground ahead at a slightly different angle, plus a colour camera and GPS. Six Pentium M computers in the back ran the software. Intel Research and the venture capital firm Mohr Davidow Ventures also supported the team.

The lasers built a 3D picture of the ground about 22 metres ahead, enough to drive safely at roughly 25 mph. To go faster, Stanley needed to see further, and that is where the camera and machine learning came in.

## How Stanley learned to see the road

The team's paper describes two clever uses of learning.

**First, learning from a human driver.** Raw laser data is noisy; small errors in the vehicle's estimated tilt can make a flat road look like a wall. To tune the many settings in its terrain-analysis software, the team had a person drive only on clear ground. Everything under the car was labelled "drivable," and strips at a fixed distance to either side were labelled "obstacle." Those labels are rough, but they are almost free to collect. Tuning on them cut the share of terrain wrongly judged as obstacles from 12.6% to 0.002%.

**Second, teaching the camera with the lasers.** In every frame, Stanley took the patch of ground just ahead that the lasers had confirmed as safe and used its colours and textures as fresh training examples. The vision system then looked for similar-looking ground much further away, out to 40 metres and beyond. When the road changed from pale gravel to darker dirt or grass, the camera model quickly adapted to the new surface.

Here is the second idea as simple steps:

1. Lasers check the nearest stretch of road: "This patch is flat and safe."
2. The software finds that patch in the camera image.
3. It records the patch's colours as "what road looks like right now."
4. It scans the rest of the image for matching colours further ahead.
5. Matching areas are marked drivable, so the car can safely go faster.

> **Try it:** Look at a photo of a country road. Put your finger on the road just in front of the camera, then trace every part of the picture that looks like that same surface. You have just done what Stanley's vision system did continuously as it drove.

## Why the win mattered

Stanley's victory turned a failed experiment into a proof of concept. In 2004, no robot finished 8 miles; in 2005, all but one of the 23 finalists passed that mark and five finished. The Stanford paper attributes Stanley's success mainly to software built on artificial intelligence, especially machine learning and probabilistic reasoning, meaning methods that work with degrees of certainty rather than fixed yes-or-no rules.

Testing discipline mattered as well. The team measured progress as the average distance between "catastrophic failures," moments when a safety driver had to take over. That figure rose from about 1 mile in December 2004 to 20 miles by July 2005, and before the qualifying event Stanley drove its last 418 test miles without one.

## What did Stanley teach the people building self-driving cars?

The biggest lesson was that a robot car should learn from data rather than depend only on rules engineers write in advance. Desert roads, lighting and dust change constantly, and no fixed rule book can anticipate them all. Stanley's habit of training its camera on the fly, using the lasers as a teacher, is a form of self-supervised learning: the machine creates its own labels from one sensor to train another.

A second lesson was to treat uncertainty honestly. Stanley's software did not simply declare a patch of ground blocked when two laser readings disagreed; it ran a statistical test that allowed for small errors in its own position. A simpler, rule-based version of the same software "saw" phantom obstacles in the middle of clear roads, while the probabilistic version did not.

The third lesson was about engineering practice: frequent desert testing, clear failure metrics and a system frozen well before race day. Those habits carry over to any team trying to make AI reliable outside the laboratory, from robot cars to medical software.

## Sources

- Stanley: The Robot that Won the DARPA Grand Challenge: http://robots.stanford.edu/papers/thrun.stanley05.pdf
- DARPA Grand Challenge (2005): https://en.wikipedia.org/w/index.php?title=DARPA_Grand_Challenge_%282005%29&oldid=970676988