Inside Job: Student Turns to GPUs to Create Drones for the Great Indoors
Marc Gyongyosi isn’t your average college student. The junior computer science major at Northwestern University’s McCormick School of Engineering has thrown himself into the world of lightweight robotics in a way that reaches far beyond the classroom.
Not only has Gyongyosi spent the past two years working with BMW’s robotics research department on developing robotic systems to help factory workers, he’s also involved in two startups. One of those, MDAR Technologies, is working on 3D vision systems for autonomous vehicles.
But it’s his work with the second company, IFM Technologies, which he founded, that landed him on a stage at our annual GPU Technology Conference.
IFM has been working on an autonomous drone that can be reliably operated indoors. Most drones today only fly outdoors because a) they’re too large and clunky to be safely flown indoors, and b) the GPS systems they rely on don’t work indoors. Further complicating the market for outdoor drones is the fact that the FAA must approve them for flight. That’s not the case with indoor drones.
Gyongyosi looked at that convergence of facts and determined that there’s a huge potential market for a commercially available indoor drone. He told GTC attendees that he estimates there are multi-billion-dollar opportunities in areas such as warehouse analytics, utility analysis, insurance inspections, and commercial real estate and construction.
And make no mistake, he’s not in this just to identify those opportunities; he wants to seize them. “We don’t want to just be a research project,” Gyongyosi said during his talk. “We want to be something that goes from problem to solution.”
His solution, however, has presented technical challenges. To start with, he’s had to find an alternative to the GPS built into outdoor drones. He said others have tried motion capture or radio beacons as GPS substitutes, but because he’s trying to keep IFM’s drone small and light, he didn’t want the extra weight. That, plus those options tend to be expensive and need constant calibration.
Similarly, other drones rely on onboard sensors to detect physical objects around them to avoid collision. But that also has presented a major space challenge on IFM’s small drone, as the amount of data that has to be processed is enormous.
“The processing power you need onboard is large,” he said. “That’s why these platforms are very large.”
To combat these issues, Gyongyosi did two things: First, he opted to mount a single camera on the IFM, sacrificing stereoscopic vision but preserving space and keeping the weight down. Then, he choose to incorporate feature tracking that operates somewhat like sensors, but instead uses the data from the camera.
When the performance of that configuration came up short of his expectations, he turned to the GPU, specifically the NVIDIA Jetson Tegra K1, which is now part of the vehicle’s physical design.
The results speak for themselves. GPUs are processing the data nearly four times as fast as a CPU. Plus, the feature-tracking rate nearly doubled, from 5.5 Hz to 9.8 Hz. And if that’s not enough, it also improved accuracy and created enough spare space that Gyongyosi was able to add a second camera, which is mounted at a 45-degree angle to the first, trading stereoscopic sight for a larger field of vision.
To further illustrate the potential impact of IFM’s design, Gyongyosi pointed to the colossal failure that is Berlin’s long-planned futuristic airport, a project that was supposed to open years ago but remains non-operational after design flaws were found in the fire detection system during inspection.
Gyongyosi believes indoor drones could have prevented the fiasco by detecting the issue long before inspection, and he hopes IFM’s drones will be performing such tasks soon.