jawfish

Darpa Grand Challenge 2005 Qualifier
john posted in technologies on October 7th, 2005

What is the Grand Challenge?
Second jawfish report.

For final results, and charts of how each team did, go to the Official DARPA site.

All pictures by Kyle Byrd-Fisher.

Watching the competitors in the DARPA Grand Challenge attempt the course at Fontana Raceway makes a few things clear. First, this robot racer thing is very tough. Tasks thats ants and bees accomplish easily are really hard to do with hardware and software. Organic brains are evolved for this and silicon ones simply are not. Second, good engineering can succeed, but not over good engineering plus money. Good engineering can’t be done without good testing and thats pretty hard when your prototype weighs two tons, can go 60 mph, and is about to be let loose without a driver.

The competitors are spending about ten days taking turns at the course DARPA has set up in Fontana California at the racetrack. Less than half will make the cut for the race in Nevada on October 8. The Fontana course is laid out using the infield road-race track, but it mostly follows a route over grass. It attempts to mimic turns on tight mountain roads, has a number of obstacles and cones to avoid, and a tunnel.

When we visited, the high speed sections saw 40 mph, but there was a lot of dead stop and slow crawl. A number of entries this day didn’t make it through the first part of the course at all, and at least one ignored part of the route.

Seeing a Jeep Cherokee get stuck in the tunnel, I asked the crew how they did previously. They said they had previously completed the course but had just changed a few parameters, obviously not for the best. The Jeep was confused by the walls, and finally couldn’t figure out how to backup and turn.

It’s important to keep in mind that the software in this race is very rough. Even if it is relatively free of mistakes, the algorithms are complex and need a lot more refinement. The only way to do that is test, try changes and test again. If a developer could just sit down and work it out on paper, robot racing would have been done long ago.

Why is the software so difficult?
The software that controls the vehicle has to look ahead and decide what the terrain is, watch the GPS and assigned route for the current location, pay attention to the attitude of the vehicle - the pitch and roll and yaw, monitor the hardware systems, decide what direction and speed to take, then do the control with steering, throttle, and braking. It has to do all this in real-time, no little clock icon with ticking hands allowed. Practically this means multiple computers running many different programs, talking to each other.

Two main types of visual sensor used by the teams are: laser radars, called LIDAR, and stereo cameras.
The image below shows two LIDARs, one with blinders.

You can see the stereo cameras in the Red Team’s pod, in the image, above.

The stereo camera software reads the camera images and by comparing differences tries to achieve depth perception. This view must be checked against the LIDAR information. One day this may be a finished system, but right now it’s at a rough prototype stage. For one thing. the data streams coming from the cameras can swamp a fast computer. The Caltech team told me they run four VGA cameras at 60 fps. Each of those streams would load a fast PC, though use of high-end video hardware can make it more efficient.

To make matters worse, the laser-radar sensors (LIDAR) aren’t really good enough. They don’t produce a complete image like a tv screen, what they return is more like one line from a sonogram. They can definitely sense small objects, like a traffic cone or a 2×4, but they can’t tell a boulder from a balloon. Some teams physically scan the LIDAR unit to get more data.

When you drive across a field, you can see a rock and tell how big it is, in 3D, and you can tell a clump of grass is soft and pliant. A shadow might hide a hole, but you avoid it from experience, or drive through because you know the area just doesn’t have holes like that. Your experience makes your interpretation of the visual data very reliable, but go scuba diving, or walking in the Amazonian forest, and you will soon be disoriented. The robots are always on the verge of disorientation. While I was watching every vehicle slowed during the fast section at the same harmless patch of greener grass. When humans are infants they crawl and touch everything around them. This joins the sensation of touch with the visual and trains the brain. These robots need the same training regimen, but there is no time, for the race is only days away.

What are the engineering trade-offs in choice of vehicles?
Some teams used cheap four-wheel drives that fit their budget, some chose a particular platform, like Carnegie-Mellon who chose H1 Hummers. Cornell uses a light-strike military car, very similar to a dune buggy. The Palos Verdes high school uses an Acura MDX because one parent had a contact with Honda.

It does seem as if many teams never considered the nature of the Mojave. Sure, they went for all wheel drive, maybe with some extra ground clearance, but many don’t have extra bumpers, skid plates, roll bars, or even stone protection for the sensors.

Here’s one proper bumper guard on #23, note the yellow feelers.

If the DARPA race course involves fire roads with a little washboard, these SUVs will be ok, but the moment it veers off into the hills and arroyos of the open desert, they are toast. It seems too obvious to point out that if you are racing across the desert, you need to start with a vehicle that a human can drive there successfully. The SCORE pickups and buggies that race in the desert have huge wheels, ultra-high ground clearance, multiple shock absorbers at each corner, and massive roll bars.

Team Ensco has a formidable buggy, seen here easily negotiating a set of turns.

The motorcycle of the Blue Team is a quixotic exercise, but there are even odder choices. Team Jefferson had four car wheels on its home-built, without off-road tires. The Terrahawk team built an articulated six-wheeler, but with a narrow track and little ground clearance, it’s not going far.

In this next sequence you see a brand-new Toyota pickup get through the tunnel, make a turn, hit a car with its unprotected stock bumper, the car rolls forward, the truck swerves around it, makes it to the hay bales, and finally drives over them, where it gets stuck. Fortunately its sensors weren’t damaged by the collision.

Less practical than the SUVs and pickups, there are several utility buggies, normally intended for work on ranches and golf courses. They are tough little vehicles, but not wide enough or with enough motor, and while very maneuverable, they don’t have the needed ground clearance either.

Here is a Polaris, note the clever use of PVC drain pipe:

That leaves the lone ATV, the mini-buggy, the home-built racer, the bizarre truck, and the true desert buggies.

The ATV has traction, maneuverability, ground clearance, and a wide stance. But like all of the smaller vehicles, it sacrifices space for sensors, computers, and electric power.

They’ve done a great job of packaging, though. here they are passing the last obstacle.

Several teams use portable 120v generators strapped to the vehicle. At least one SUV had a rank of four computers and UPSes, plus sensors, network switches, and control motors. That’s a lot of electric power. At the opposite end of the spectrum is the TerraMax truck, an all-wheel drive, rear-steering monster built for the military. It did well in the qualifying, but won’t be able to handle steep gullies or tight switchbacks. It is intended, no doubt, for military convoy duty, where it can drive in a line like elephants at the circus.

The street vehicles do have an advantage over the desert racers: they have fly-by-wire components already- with power steering, ABS brakes, and cruise control. Still, its easier to retrofit those items to a racer, than turn a car into something with 24″ of clearance.

My analog hardware performance bets are on the Hummers, the Cornell striker, the buggies, and a long shot with the ATV. The rest of these designers have clearly never seen a desert race.

More on the race to come.