Since 2016, the human pilots of the Drone Racing League have competed to see who could whip a quadcopter around pylons and through hoops the fastest. On Tuesday, they’ll get a new challenge: the fully autonomous RacerAI, a drone programmed to fly itself.
Nine teams of programmers from around the world have have been coding for months to come up with the best software to control the Drone Racing League-designed RacerAI. Their work, along with the drone itself, will debut at the Addition Financial Arena in Orlando, Florida. The software needs to take advantage of the drone’s four cameras, four propellers and Nvidia processor.
The RacerAI looks more like a flying fish or bird of prey than a conventional quadcopter. Its arrow-like body emphasizes its purpose: full speed ahead.
The race will be the first in the new Artificial Intelligence Robotic Racing series, which uses simpler courses that pit computer-piloted drones against each other. The series culminates on Dec. 6, when the best AI-piloted drone will take on a human pilot in Austin, Texas.
“We’re here to watch as robotics evolve beyond humanity,” said Ryan Gury, DRL’s chief technology officer and designer of the league’s current racing drones. “We believe in the future of autonomous robotics.”
Competitions that pit humans against machines can be compelling. IBM computers famouslyand , and kicked up the difficulty level with its wins against the best players of the Go board game. We all know computers have us beat when it comes to doing math and remembering anniversaries, but it’s somewhere between fun and scary to watch the bots win in other domains.
The physical world is a less cerebral AI challenge than translating French into English or self-driving cars, passenger aircraft or delivery drones that have to deal with their surroundings.. Boston Dynamics’ humanoid and its . Similarly, the DRL’s RacerAI is designed to physically navigate on its own in the real world. Granted, it’s the limited domain of a race course inside a big arena, but it isn’t hard to see how this technology can apply to
Nicholas Horbaczewski, DRL’s chief executive and founder, expects the drone competitions will affect sports and more.
“Drone racing…sits on a blurry line between the digital and real, which makes it the perfect testing ground to accelerate AI innovation and expedite head-to-head machine vs. human performance,” he said in a statement. “I believe you’ll see a rise in robotic sports as well as an expanded use of autonomous drones, transforming an array of spaces from emergency response and aerial surveying to mobility and urban package delivery.”
Gury is a see what the drone sees over a wireless link. But he thinks AI-powered drones will ultimately prove superior. When? “2023 is our bet,” he said. “Everything really begins to shape up when you see robots outperform humans physically.”himself, one of those people who dons a headset to
RacerAI drone design
When human pilots fly a racing drone, a radio link lets them see things from the drone’s perspective. With the RacerAI, all the thinking takes place on the drone itself — specifically on an Nvidia Xavier processor designed for autonomous vehicles.
It’s pretty power-hungry, consuming 40 watts of power. That’s something like 20 times the power your phone processor uses. The RacerAI can fly only 2 to 3 minutes on one battery charge, about the same as the human-piloted racing drones, according to Gury.
Most drones, including the DRL’s Racer4 that the league’s human pilots fly and its record-setting RacerX that hit a top speed of 179.3 miles per hour, are an X shape, with two propellers in front and two propellers in the rear. The RacerAI takes a different approach — the shape of a plus sign. The propellers on the front, left and right point downward, while the rear propeller points upward.
Diagonal struts connect the front propeller to the left and right propellers, and each of those diagonals has a pair of fisheye cameras about 8 inches (20cm) apart. Each camera pair can be used to see in 3D stereoscopically, like human eyes. With two pairs, the drone gets that 3D vision ability for the entire 180-degree view, Gury said.
Nine drone programming teams
The nine AIRR teams, selected from more than 420 that tried out, are from around the world. They have names like the Warsaw MIMotaurs from Poland, MAVlab from the Netherlands, Team Puffin from Sweden and Icarus from Georgia Tech in the US.
Their job is to write software that interprets data from the cameras and the drone’s inertial tracking system and then instructs the drone on how to fly.
This year, the autonomous drone race courses won’t be as complex as those the human pilots tackle. Expect straight lines and basic slaloming, with gates helpfully marked so they’re easy for computer vision systems to spot. The first drone races will likely each take about a half minute.
But DRL will make the courses harder. “As we see competition evolve, we start to raise the stakes,” Gury said.
Bot versus bot
Researchers can learn a lot by pitting one chess-playing computer against another, and indeed the AlphaGo system partly learned how to win by playing itself. But the rise of robot pilots raises a new question for the relatively young league: Will people want to watch machines races machines?
It’s easy to anthropomorphize the physical ordeal that BattleBots face as they smash, pound and saw each other. But even there, a human is behind the remote control. It’s harder to put yourself in the shoes of a programming team trying to optimize training data and figure out how many layers deep a drone’s neural network programming should be.
It’s definitely enough to get Gury’s juices flowing: “What we want to see is the greatest autonomous drones in the world.”
First published October 8.
Update, October 9: Adds comment from DRL CEO.