Yesterday, DARPA held a media briefing detailing what we have to look forward to at the DARPA Robotics Challenge Finals next month at Fairplex in Pomona, Calif., on June 5 and 6. We’re absolutely convinced that this is going to be one of the most exciting robotics events of the decade, so let’s take a look at what’s going to happen.
DARPA program manager Gill Pratt provided an overview of the event, along with some logistical details for people who can’t make it out to Pomona, as well as for those who can. The briefing was an hour long, but we won’t subject you to all of it, since there was a lot of background that we’re already familiar with.
Here’s what we found most interesting; all of the quotes are from Dr. Pratt:
“When you look at the DARPA Robotics Challenge, you’ll see a bunch of robots that are very impressive mechanically. Some of them may look like an imitation of a person, some may look like an imitation of some four-legged creature… There’s all different shapes and sizes. But you’re actually not seeing the most important part of the technology we’re trying to improve, which is really, how do human beings and robots work together when they’re separated by a significant physical distance, and the communication link between the two sides is severely degraded.”
This is a crucial thing to keep in mind throughout the DRC Finals: the robots themselves are awesome and fun to watch, but the really difficult problem is in the control of the robots, in the software and user interface.
This is a characteristic of robots in general: they’re tricky to physically construct, for sure, but getting them to reliably do anything once you’ve put them together is extraordinarily difficult, especially if you mix in real-world issues like communication limitations. And once we’ve managed to get the DRC robots doing reliably useful stuff in a timely manner (which is what we hope to see at the Finals, and will likely involve a significant amount of both supervised and unsupervised autonomously), we can adapt what we learn to other hardware platforms that have a shot at performing real-world tasks.
“There’ll be 25 teams; a substantial fraction of them will have difficulty during the challenge. And we do that on purpose: DARPA takes high risks for high rewards, and so the challenge is quite hard. We don’t allow any recharging of robots during the event. There will be no hoists or safety ropes. If a robot falls, and we imagine that a large number of them will, it will face the danger of being injured, and perhaps not being able to continue. If it does fall, it will have to get up on its own.
If a team gets stuck, and [the robot] cannot get up on its own, they have the choice to elect to take a 10-minute penalty, and we’ll do a simulation of what it would be like to send in a second robot in order to continue the mission after the first one has gotten stuck. The simulation of that will be done by the team retrieving the first robot, bringing it back to the starting line, and if it’s still in working condition, starting it over with a 10-minute penalty.”
You should expect to see a lot of robots falling down, a lot of failures of various sorts. We are expecting, hoping at least, that some of the best teams will manage to do most, if not all of the tasks, and we think that some of the teams will in fact do these tasks without falling down.”
Yup. Expect to see lots of falls. But, expect to see far fewer falls than we did in the DRC Trials, both because teams have had a lot of time to practice since the Trials, and because falling is a much bigger deal this time around.
Since teams only get an hour to complete eight tasks, falling and not being able to promptly recover almost certainly means missing a task or two, as a reset automatically knocks out a sixth of the time, plus however long it takes the robot to get back to where it fell.
Keep in mind that when Dr. Pratt says that “we think that some of the teams will in fact do these tasks without falling down,” he’s well aware that some of the teams are fielding robots that aren’t bipeds and are much more stable than most of the humanoid designs like ATLAS.
“In addition to the robot competition itself, we have an exposition that is going to show all sorts of neat things; the history of DARPA, what the agency has been about, and some of the great work that we’ve done before. We’re also going to show off a lot of robotic technology from different companies and universities, as well as first response technology. There will be a very large demonstration area, and we’ll have a number of dynamic demonstrations of different robots and robots for disaster response.”
The DRC Finals are totally free and open to the public: you can just walk right into Fairplex on June 5 and 6 and see all the action. There will also be a huge expo, but it’s a lot more than just DARPA: all kinds of robotics hardware and software companies will be exhibiting, and it’s going to be amazing. If you’re anywhere near Los Angeles, find a way to get there, and if you’re not, the Finals will all be webcast in realtime HD on YouTube (we’ll have more details on that closer to the event).
There’s also going to be a workshop on June 7, immediately after the competition, featuring a description from the three winning teams on all of the strategies and techniques that they used to come out on top. This won’t be open to the public, but it will be webcast, and media will be allowed in, so we’ll be there taking notes.
We were able to ask Dr. Pratt about what to look forward to with the tasks, although he wasn’t giving out a lot of details. That’s part of the point, though: DARPA doesn’t want teams to know too much, so that they’ll have to adapt a bit to each task. This is a significant departure from the Trials, where every single detail about each task was known in advance.
“The tasks that you’re going to see that are assembled into a mission at the Finals are similar to the ones that you saw at the Trials, but they aren’t exactly the same. We’ve been very purposefully not giving blueprints out to any of the teams that say, this is exactly what the door handle will be like, or what the valve will be like. We want to push the teams far away from trying to have scripted responses or a very close idea of what they’re going to face, and the surprise task is of course the biggest part of that.
At the same time, we don’t want to make the problem too hard so that they’ll just fail, and so we’ve had a beta test event a little over a month ago where most of the teams came and got to see what the tasks would be like. But again, we made no promises saying that it was exactly what they were going to be like. The teams will arrive a few days before the contest itself, so they’ll be able to see the course, and there’ll even be a walkthrough of the course. But we also told them that we may change the course from day to day, and to push them again away from overfitting the solution to the particulars. And so in general, the tasks are going to be very much like what you saw the last time, but we’re going to change them just enough so the teams can’t just hit the play button and expect that to work.”
There’s going to be a media preview day on June 4, which will be our first look at the course, so as soon as we get back from that, we’ll be able to tell you what the Finals will look like.
Now, if you’ve read this blog for very long, we hope that you’ve got realistic expectations when it comes to the current state of robotics. Having said that, we’re quite honestly expecting to be impressed by the progress that the DRC teams have made since the trials. This is expectation that’s severely tempered by reality, of course: like, if a bipedal robot manages to complete all eight tasks without falling over in under 60 minutes, we’re going to be damn impressed.
Even if that doesn’t happen, things are going to be much more impressive than we’ve seen before, as teams imbue their robots with significantly more task-level autonomy, allowing them to move much faster and more confidently:
“We had talked before, during the trials, that the mobility and dexterity of the machines was about the same as a one year old. It’s a little bit difficult to make an easy analogy for the finals. The level of autonomy going on now is at the path level, which means a team (some of the better teams) can say, ‘open the door,’ and the machine can figure out where the door is, and do that task on its own. The machine can’t figure out, ‘hey I tried to open the door but then I got stuck so now I have to do something different.’ That strategic thinking is still not happening in the machine itself.
What the teams have done is moved a lot of the low-level mobility and manipulation skills from something that needed to be micromanaged by the human operators to the machine itself. At the task level, I think that some of the better teams are there.”
Whatever happens, the DRC Finals represent the cutting edge of humanoid robotics. No robots have ever been as complicated or as capable as the ones that we’re going to see compete in June. And although we don’t have the robots that science fiction has been promising us for the last several decades, the real robots are way cooler than science fiction robots could ever be, because they exist:
“There’s a huge difference between science fact and science fiction. We all know what robots in science fiction are like; robots in science fact are a little bit more modest because the technology isn’t anywhere near what science fiction would have us believe. But, they’re also more exciting, because they’re actually real and they point the way to what may be possible in the future.”
“These are prototypes used to advance technology that perhaps, 10 years from now, might be seen in devices that are actually out in the field.”
And 10 years from now, we’ll be covering those devices, too.