kwc.org Photos Spare Cycles MythBusters

Talk: Winning the DARPA Grand Challenge

I went to go see Sebastian Thrun speak at Stanford about his team's winning effort in the DARPA Grand Challenge. Thrun described the contest as how to stay on the road for a very long time. It was not a general path-finding problem: DARPA gives you the route with a corridor you have to follow, as well as speed limits you have to observe for various parts of the course. Of course, DARPA didn't give perfect data. He showed a video generated from the data of Stanley driving through the most dangerous part of the course: the switchbacks of Beer Bottle Pass with a cliff on one side. DARPA's corridor was overlaid on top of Stanley's sensor data and it was easy to see that much of DARPA's corridor was actually over the cliff.

During training many traffic cones were "frequent victims of computer glitches," but Team Stanley was called "Team Boring" by the cleanup crew for their lack of incidents. The actual challenge was described as getting the data at 4AM, getting Stanley to the starting line at 6AM, and then sitting around drinking beer for several hours. The big moment came when Stanley passed CMU's Highlander as the CMU and Stanford teams listened to race radio. Thrun narrated the exciting finish for us: "[The head of DARPA] is waving his flag as if the car could see it."

Thrun said that they won mostly through luck given how close four of the teams finished. The speed limits set by DARPA for the various parts of the course were too conservative, so the cars were running below their full potential. DARPA also decided to make the course fairly easy. Asked if CMU would have won had they not had engine problems, Thrun answered, "In all likelihood, yes." Also, Team ENSCO had a faster average course time but flatted on "something really big CMU left behind" (the CMU part may have been a joke). Thrun felt that Stanford had better software than CMU and on a tougher course Stanford would have the advantage.

In the future, Thrun wants to try driving 65mph on 280, parking in a garage, convoy driving, and driving assist. Part of his motivation is to reduce traffic deaths, which a driving assist system could help prevent. He also feels that a fully automated system would change society by allowing you to use your commute time productively -- you could even drive to your destination, get out, and then send your car to go park in a parking garage farther away. These are still looking far ahead. In response to someone asking what it would take to drive at human-controlled speeds, Thrun related it to asking the Wright brothers, "If you want to fly over the Atlantic, what's missing?"

More on CMU

The main difference between Stanford and CMU's system was that CMU did a lot of preplanning. They used a lot of people to look over the course, overlay imagery, and try to mark potential hazards for the vehicle. Stanford did no preplanning and instead had a simple velocity control with two parameters: shock limit and speed recovery rate. If Stanley encountered a shock, it would slow down to it's shock limit and then ramp back up to it's desired speed. Thrun noted that it generally better to slow down before the shock (e.g. a bump), but it's a good idea to slow down after one as well as the shocks often come in bunches.

Sensors

They used lasers and a non-stereo vision for sensing. The lasers were angled downward to scan a plane ~25m in front of the vehicle. If there is a discontinuity along the plane of the laser they know there is an obstacle, though it's not that simple. They also have to find obstacles like berms, which generate no discontinuity. They integrated the laser data over time to try the berms, but this also created problems. As the vehicle bounces up and down, the lasers may scan the same point several times and general false positives for obstacles. They used a probabilistic model of their measuring error, with machine learning used to set the parameters in order to eliminate a lot of the false positives from the bouncing. The combined techniques give them a mean time to error of about 20 miles.

The lasers alone can only get Stanley up to 25mph as they only scan 25m in front of the vehicle. In order to get up to 35mph they have to incorporate vision to see further up the road. Although they bought the best stereo vision sensors on the market, they went with non-stereo techniques as Thrun felt that stereo was a bad idea -- small errors in the angles of the camera would cause huge measuring errors and not even humans use stereo vision for this purpose.

They fed the laser data right into the vision data. They laser data would project a trapezoid plane of where the road in front of the car was and the vision system would use this data to learn where the road further ahead was. It seemed to be largely based on color. If the car drove onto the grass, the grass would become 'road.' If the car drove into the tunnel, "even the walls become drivable." This wasn't a problem as the vision system was used for warning, not steering.

The main limitation on going faster than 35mph is perception. Motion blur, lighting, etc... make it difficult to for the vehicle to see far enough ahead to drive at higher speeds.

Planner

They rolled out forward trajectories and picked the center of the good trajectories. There was some additional adjustment using an algorithm that tried to keep the vehicle on the center of the road. They only had to plan out forward trajectories as Stanley (unlike TerraMax) was incapable of backing up.

Post a comment


tags.

what is this?

This page contains a single entry from kwc blog posted on October 31, 2005 9:28 PM.

The previous post was Photos: 99 Islands.

The next post is Hands on iPod with video, mixed impressions.

Current entries can be found on the main page.