Iris Automation’s ‘Casia’ onboard collision avoidance system. (Iris Automation)
Iris Automation, a Silicon Valley-based startup, has been working on an onboard collision avoidance system that — for the first time — will power an FAA-sanctioned flight beyond an operator’s visual line of sight (BVLOS) without assistance from visual observers (VOs) or expensive ground-based radar systems.
Currently, almost all drone operations sanctioned by the FAA are prohibited from BVLOS operations, and the few counterexamples require either VOs or radar — both of which are prohibitive to conducting scalable, cost-effective operations such as pipeline inspection, drone delivery or urban air mobility.
A drone, the regulatory agency argues, needs to be able to sense and react to unexpected objects in the airspace just as a pilot would, and that typically means the remote operator has to act as its “eyes” and react as its pilot.
But Iris’ onboard collision-avoidance system, called Casia, has now been approved twice by the FAA to fulfill that requirement. On July 31, as part of the FAA’s UAS Integration Pilot Program (IPP), the University of Alaska successfully flew a 3.87-mile pipeline inspection mission powered by Casia, in coordination with eight Echodyne ground-based airspace management radars.
“The ability to fly BVLOS missions without ground-based radar or visual observers is a significant advancement,” said Mike Kelly, senior UAS coordinator for Westar Energy, one of the companies participating in the Kansas IPP. “Being able to operate under this waiver allows the Kansas IPP team the ability to research and develop truly scalable BVLOS UAS operations for the automated inspection of linear infrastructure.”
And the Kansas Department of Transportation (KDOT) has received permission from the FAA to break the next barrier. In the next few weeks, KDOT will conduct a nine-mile BVLOS drone operation to inspect transmission lines leveraging only the Casia onboard detect-and-avoid systems. No visual observers, and no expensive ground-based radar.
“You could do a 300-mile mission without setting anything up,” Iris CEO and co-founder Alexander Harmsen told Avionics. You don’t even have to visit those 300 miles before to be sure you’re going to be able to avoid things as they come up. It’s the same thing as pilots – the FAA trusts me to fly in new areas where I’ve never flown before because they believe that I can use my eyes to avoid intruders when I get there.”
A wide array of companies, as well as NASA and the FAA, are tackling the problem of how to plug drones into the national airspace, ensuring that operators can efficiently communicate with air traffic controllers to avoid manned air traffic and receive flight plan approvals and airspace alerts. But avoiding unexpected obstacles in the airspace — “That’s the final nail…that’s the biggest piece,” said Harmsen.
“Just like pilots right now or when I fly, when I’m a [pilot in command], I’m the one that needs to avoid crashing into anything, right?” Harmsen said. “Even when I’m in controlled airspace, ATC is guiding me around. If there’s something that I think is going to cause me harm, then I need to get out of the way. I need to take that last second avoidance maneuver, if the comms cut out, I need to be able to avoid other aircraft and it’s going to be the same thing with drones.”
That’s what Casia is for. Using an off-the-shelf camera and processor, the system detects and classifies far away objects more accurately than a human and then determines what action the drone should take, choosing from a ‘dictionary’ of approved maneuvers. The latter part has been critical to proving the system’s safety to regulators, Harmsen said. The entire system weighs 300 grams and costs less than $10,000, according to the CEO.
|Want more eVTOL and air taxi news? Sign up for our brand new e-letter, “The Skyport,” where every other week you’ll find the most important analysis and insider scoops from the urban air mobility world.
“We went out a couple of years ago, we built this amazing algorithm that would take all of our 3D trajectory information and then do path planning around it, trying to estimate the best possible move at any one point,” Harmsen said.
But in talking with the FAA and customers, Harmsen found that what was needed was not necessarily an autonomous system, but rather “some sort of automated maneuver that’s pre-chosen,” he said.
For regulatory agencies, it’s much more difficult to confidently certify a system that could take an unlimited number of actions in any given situation. But with a deterministic output like what Iris has built, “you can say, we have five different menu options, we’ve tested each one a thousand times on four different aircraft types, and it works,” Harmsen said. “We know the mean time between failure is when we run this in simulation.”
It’s a similar system to what Amazon presented in its FAA filing, requesting waivers and approval to carry out Prime Air drone delivery as a commercial carrier under Part 135. Amazon included a chart describing the autonomous features its sense-and-avoid system is capable of, along with what sensory inputs and situations map to each feature. With Iris’ system, Harmsen added, operators can choose ahead of time which maneuvers are allowed and modify components of them, like how much a drone will descend in a particular maneuver.
Before reaching the deterministic outputs, however, the Casia system uses a mix of computer vision and artificial intelligence to understand its environment, mixing non-deterministic “thought” with deterministic outputs in a way that allows regulators to understand and certify how the system works.
Casia demonstrating single-target identification. (Iris Automation)
“From when we started building the technology, we were very intentional about not just having one big AI algorithm, one end-to-end deep neural net that controls the whole thing,” Harmsen explained. “We specifically broke it down and we have a mix of geometric, deterministic algorithms that rely on pixels and tracking — sort of traditional computer vision — together with more machine learning, AI to be able to add semantic understanding [and] things like classification systems.”
“So, very intentionally there’s a mix of those two things; it’s a mix of making the system generalized, making it interpretable so you can actually understand what’s happening at each stage of the process and making sure that it can actually be regulated because the FAA and different aviation authorities around the world, they want to be able to pull it apart and understand what’s happening under the hood,” Harmsen said. “And so we’re very intentional about that.”
That design has led to the two groundbreaking FAA-approved flights, which could enable cost-effective commercial drone operations — and potentially more. Ultimately, drones and UAM aircraft may have onboard radars installed, expanding the sensory inputs and therefore the capabilities of collision-avoidance systems like Casia.
“Like we pull in UTM and ADS-B, it would be great to pull in this radar information, especially if it’s onboard radar,” Harmsen said. “And then we can offer this sensor fusion package similar to self-driving cars. The more sensors, the more feeds you feed in, the better it is.”
“There are a couple companies that are starting to do that,” Harmsen added. “The biggest problem is shrinking these radars down to a form factor that can be put on these drones and be useful; it’s expensive, and it’s hardware development so it’s very slow … But down the road, especially when we get to air taxis, larger cargo-type delivery drones, we hope that the radar companies will come down in cost, size, weight and power so that we can do sensor fusion with them as well.”
Iris is working with a number of air taxi companies to test Casia’s capabilities within that mission set, Harmsen confirmed, though he couldn’t publicly disclose which companies. There may defense applications for the technology as well, though Iris said it is more focused on commercial and industry uses for the moment.
First, however, Casia has to prove itself in the upcoming trial with KDOT.
This article has been updated to reflect additional comments from Iris Automation.