Back in May, I mentioned that Off Wing reader Dave Smith and some colleagues from Mitre were taking part in the DARPA Grand Challenge, a competition where DARPA is challenging companies to build an autonomous vehicle that can travel unassisted for 175 miles.
Well, the first segment of the Grand Challenge was completed yesterday at California Speedway in Fontana, and Dave has been kind enough to send in some reports from the field. Because this is a bit off-topic, you can find Dave's notes to me after the jump.
Wednesday, October 5, 12:07 p.m. U.S. PDT
The MITRE team is one of 43 teams here at the California Speedway for the National Qualifying Event for the Grand Challenge race. As far as we know, only 20 teams will be allowed to enter the race for $2 million. The race route has yet to be announced, but will begin and end in Primm, Nevada this year, and will cover 175 miles.
The 43 vehicles that have made it to this point span a huge range, from a 350 pound motorcycle, several ATV or utility vehicle based machines, quite a few pickup truck or SUV-based vehicles like our own, all the way to a giant 15-ton Oshkosh truck, which barely fits through the 10 foot wide by 9 foot high minimum course dimensions.
DARPA has several 3-mile courses laid out in the infield of the race track, with lots of obstacles such as metal gates, telephone poles, a tunnel, concrete barriers, "rough road" strewn with 4x4's and tires, sharp curves, hills and side-slopes, vehicles parked in the middle of the course, and a "tank trap" in the middle of the course right before the finish line.
We were told that we'd get three or more attempts at the qualifying runs, where we're scored on time to complete, number of waypoints achieved or missed, and number of obstacles avoided or struck.
Our first run last week ended very quickly; as our truck approached the first obstacle, a gate composed of a pair of metal barriers set 10 feet apart, it declined to go through, and started heading around it.
Unfortunately, it got a little too close to the concrete barriers that separated the qualifying course from the spectators, and DARPA hit the remote E-stop button, ending our run. By looking at the log data from the run, we learned that we had set our safety cushion, which keeps us from getting too close to obstacles, too conservatively, and from the angle at which the truck approached the gate, with traffic cones presenting additional obstacles nearby, it didn't appear that there was enough room to go through safely. We believe that our robot would have safely gone around the gate and completed more of the course if DARPA hadn't pulled the plug... but our first order of business was to make sure that we'd make it through the gate next time.
I made a quick run to Home Depot for corrugated steel, 2x4's, and shelf brackets, and in a couple of hours we had our own pair of barriers to use as a gate in the practice sessions we're allotted between qualifying runs. After tweaking a few parameters (and folding back the mirrors on the truck), we were able to repeatably get through a gate just like the one on the course, with traffic cones all around, from any angle. We felt pretty confident before our second qualifying run.
However, as we prepared for our run, we watched the team that ran before us, "CIMAR" from the University of Florida, go haywire right in the middle of what appeared to be a promising run. Apparently, they lost their GPS lock, which meant that their robot had no idea where it was.
Apparently ESPN showed a clip of their vehicle driving around like a chicken with its head cut off. As soon as DARPA pulled the plug on their run, they started our vehicle... and apparently, we had lost our GPS lock as well, and immediately drove to the right of the first gate and toward the spectators again! DARPA stopped us right away, and we were a very disappointing 0 for 2.
Our best guess is that there was a momentary problem with the GPS satellite constellation just at the wrong time for CIMAR and for us. We had a plan for addressing loss of GPS signal by reverting to our inertial navigation system, but never implemented it, as we'd never encountered the situation before. Some of the teams have inertial systems that cost over $100,000 - more than our entire vehicle, computer, and sensor package - but ours should be good enough to survive a brief GPS outage. Still, at this late juncture, we decided not to make any major changes to our software, and just hope for the best in our remaining qualifying runs.
On our third run, we made it through the first gate, about 1/3 of the way through the whole course, and into a tunnel, which we've successfully navigated many times back in our July testing in Nevada. However, this time, we bumped into the jersey barriers at the side of the tunnel, and couldn't continue.
It turned out that this tunnel was a bit wider than tunnels we've negotiated before. When we enter a narrow tunnel and lose our GPS signal, we immediately use a reactive obstacle-avoidance behavior to stay away from the walls of the tunnel and keep moving to the end of the tunnel, where we expect to resume GPS navigation. However, because this tunnel was a bit wider, we didn't switch to reactive mode right away, and by the time we did, we were moving a little too quickly toward the tunnel wall to avoid it in time. Our vehicle escaped with a few scratches on the bumper, but we were again very disappointed with the performance.
I made another run to Home Depot, and within a few hours, the whole team had helped construct 14 4x8 "tunnel walls" from home insulation sheets, picket fence slats, zip ties, and shelf brackets. We practiced entering a wider "tunnel" by covering my baseball cap with aluminum foil and using it to cover our GPS antenna just as the truck got between the 56-foot walls we had placed 14 feet apart in the practice area. After a few more parameter tweaks, we were confident that we'd be able to swish
the tunnel every time.
It was only Sunday, it looked as though we'd still get another one or two qualifying runs, and at that point only about 18 teams had made it all the way through the course at least once, so we still felt hopeful that we could qualify to be one of the top 20 teams.
Our next qualifying run was almost completely successful! We made it through the metal gate, through the tunnel, through a slalom and "rough road" area, and past the parked cars and tank trap, without touching any of the obstacles, though we did swing a bit wide of a few of the cones marking the (otherwise invisible) course boundaries along the way. That made us one of about 23 teams to have completed the course to that point.
Monday evening, DARPA announced that 10 teams (which had successfully navigated the course three or four times already) had been selected to enter the Grand Challenge race, 15 teams (which had never completed the course) had been eliminated, and 18 teams were still in contention and would have another qualifying run on Tuesday. We were in this group on the bubble, which is about all we could've hoped for given our performance to that point.
On Tuesday morning, we had another successful run, with the exception of very slightly brushing one of the parked cars near the end of the course. Around noon, DARPA informed us that we were now among 7 teams that were *still* on the bubble. We assumed that about 7 of the 18 previous "bubble" teams had made it to the big event based on their performance in the morning, a few had been eliminated, and that this new 7 were now competing for about three remaining slots - though we began to hope that DARPA would allow more than 20 teams into the race. (Originally there were only supposed to be 40 teams here at the qualifying event, but they had room for 43, so they allowed a few more in. Will they do the same for the race?)
We had one final run Tuesday afternoon, and it was our best run yet. We believe that another couple of teams eliminated themselves with DNF's. Our best guess is that we're now ranked either #20 or #21 - but DARPA has not revealed their scoring formula, so there's no way to know for sure. Our times have been slower than most of the others, but we've been reluctant to speed things up because our first priority has been finishing cleanly. Will that count against us? We don't know. We are awaiting the official announcement of the list of teams that have made it to the Grand Challenge race at the closing ceremonies this afternoon.
Regardless of how that goes, we're pretty happy with how we've performed out here - given that 19 of these 43 teams were in the race last year and had at least an 18-month head start on us and a chance to practice the qualifying course last year, we knew it would be very tough to break into the top group.
I'll let everyone know what we hear this afternoon, but at this point it's out of our hands. You may find out as soon as i do by checking DARPA's website, http://www.grandchallenge.org/. There are also lots of pictures and scoring results from the qualifying runs there. I'll have lots of pictures to send later, but for now, i've gotta go....
...dave
Wednesday, October 5, 4:43 p.m. U.S. PDT
Well, the 43 teams gathered today at the closing ceremonies for the DARPA Grand Challenge National Qualifying Event, and Tony Tether, director of DARPA (who got some notoriety when he proposed the misunderstood "Total Information Awareness" project) announced that unfortunately, only 20 teams would be selected to compete in the Grand Challenge race on Saturday.
We all waited as he listed the names of the teams that had been selected, in no particular order. Our hearts sank when he had listed 20 teams, and we were not among them! But then one of the deputy directors whispered something to him, and they acted out a mock debate over whether they should allow a few more teams to enter. It was all an act to increase the suspense, and they "decided" to allow 23 teams in! They named the final three teams, and we are in!!!
Now that we've made the few tweaks necessary to pass the NQE, the slate is clean and we feel good about our chances to be competitive in the race. We may not be the fastest, but we're hoping that slow and steady will win in the end. Today we pack up all of our equipment for the trip to Primm Nevada, and tonight we celebrate.
...dave
Wednesday, October 5 8:36 p.m. U.S. PDT
For people who are interested in more than just MITRE's entry, i can add that there has been an interesting development among the favorites. Last year, Carnegie-Mellon's "Red Team" had one entry, "Sandstorm". This year, their extremely well-funded team features a second entry, "H1ghlander". (Their army of 50 team members hangs out in one of Team Target Chip Ganassi's support trucks when they're not in the paddock!)
Both robots are based on Hummer H1's. This year, DARPA has only announced the seeding of the top three teams. H1ghlander has the pole position, and Sandstorm will depart third. But from what i've seen, the team that will be starting second has the best chance to win. Stanford's "Stanley" is heavily funded by Volkswagen and Red Bull. It's a VW Touareg painted just like VW's Paris-Dakar entry and other racing teams sponsored by VW and Red Bull. They even have an identical backup race vehicle, as well as a fleet of Tourareg's to transport the team members. They negotiated the NQE course quickly, smoothly, and flawlessly every time, as far as i could see.
Before seeing the competition, i didn't think anyone would win the Grand Challenge prize this year, but now... i'm not so sure.
We shall see!
...dave
