Daedalean & The AI That Knows Where Not To Land
As someone who has recently became an owner of a fairly light and harmless drone, I know that almost everywhere in the world laws about it are absolutely bonkers. In my home country of the Netherlands, I can’t even fly it in my own backyard. Ironically, the King of the Netherlands, Willem-Alexander, may or may not have even gotten a fine for buying a small drone and flying it in his own backyard because the law states that no one may fly above property of Dutch royalty.
While some countries are less strict, the laws usually say something like this: no flying above crowds, no flying after dark, no flying beyond your field of vision, no flying closer than 100 meters to any “structure,” and no flying higher than 100 meters. This is because some very dumb people bought drones and did some really dangerous things with them. Today this technology has become a lot safer by adding obstacle avoidance sensors and using GPS to not allow drones to launch in or fly into restricted zones. Right now, if a drone needs to make an emergency landing, it reaches a high altitude to avoid objects and then descends and lands on the exact spot it took off from even if that means landing right on top of your head. Thanks to Daedalean’s AI, that might soon be an issue of the past.
Magpie, Proof of Concept for Drones
Starting today, Daedalean and UAVenture are announcing Magpie, a light entry version of their AI that uses one simple downward-facing camera and neural-network-processed computer vision than will be released before the end of the year. This enables GPS-independent navigation and allows drones to find safe landing spots, but the real magic is in how it works, which we will explain later in the article.
Magpie is intended for professional drones and will enable them to perform autonomous tasks, avoid obstacles, return home, and find a safe landing spot. This kind of drone oftentimes doesn’t require certification, but in some situations, like when a drone is flying over inhabited areas, using Magpie can actually help get the approval needed since it “removes or reduces risks to people, property and critical infrastructure.”
UAVenture already has an autopilot for drones called AirRails that can perform somewhat similar functions but is based on radar technology and is thus limited. Magpie will be available as an upgrade to AirRails, and because Daedalean’s end goals are aimed at large eVTOL vehicles rather than drones, distribution will be totally in the hands of UAVenture. It’s a great partnership of convenience that gave AirRails a helpful upgrade and gave Daedalean a proof of concept and a lot of data from test flights that UAVenture helped enable. The data is crucial for their next steps in creating an AI for helicopters and eVTOL vehicles.
There are hundreds of eVTOL startups, but basically only a few companies like Daedalean that will actually make them autonomous and open the market for mass production, so let me tell you a bit more about them and their crucial role in the industry.
Daedalean — AI & Certification are the Secret Sauce
As mentioned before, Daedalean AI is an industry-leading autonomous flight control software development company. It is based in Zurich, Switzerland. Cofounded by CEO Luuk van Dijk and CPO Anna Chernova, who met while working in nearby offices at Google’s headquarters, the team has grown to more than 30 people. We got the chance to talk to Anna Chernova at length about a wide variety of interesting subjects, but let’s start with the story of how Daedalean was founded.
At the time, the biggest basic issues with adding some automation, like flight stabilization to toy helicopters (now commonly known as a drone), was basically solved, and due to her interest and experience with flying real helicopters, Anna wanted to see this applied to helicopters and VTOL vehicles — as well as automate it all. Initially, she had trouble finding people willing to work on this because many believed there was no clear path for certification within the next couple of decades. Luuk, on the other hand, during his time working for SpaceX, observed how it could be done. Software engineers at SpaceX are forced to write their code in specific ways that conform to the parameters that NASA requires for certification. Without doing that, the dragon capsule would not have been permitted to dock with the International Space Station.
Considering Daedalean’s successful innovation partnership with the EU Aviation Safety Agency (EASA), this method towards certification of neural networks turned out to be a very realistic approach. There are hundreds of startups working to make eVTOL vehicles a reality, and most of them turn to and work with Daedalean for automation, AI, and certification. Daedalean’s success has enabled the startup to raise a total of $12 million that it needs in order to continue operations, expand, and develop its AI. While at first glance that may not sound like a lot, if you consider the fact that the product mostly consists of programming and not the engineering of expensive, large, dangerous flying machines, that’s a lot of money. Some of the greatest challenges Daedalean faces are creating standards for AI transparency as well as robustness and certification for the neural networks behind it. Not only should the AI not fail while transporting people in the air, but it should also be possible to find out why the AI made any and all decisions.
Under the Asimolar Ai Principles, generally known as the 23 principles designed to keep AI safe and make sure it works for the common good, safety and failure transparency happen to be the two highest ethics and values principles on the list. Just to refresh your memory, the Asimolar AI Principles is a list of principles that has over 4,000 signatories, including very notable ones like Stephen Hawking and Elon Musk.
So, How Will They Do It?
According to Daedalean, the best way to teach a drone to fly is to give it a piloting exam. Apparently, the final examination for humans has a long but well defined list of tasks that a pilot needs to be able to understand and do. Also, a bit more than a third of those tasks involve vision. Daedalean’s approach is to teach its AI new tricks, slowly tick off all the boxes on the checklist, and come up with tests that could be used to certify specific functions and capabilities of autonomous flight AI.
Not only are the company leading in this field, it’s creating a path that other companies can use to follow in its footsteps to success. This path starts by teaching AI how to create visual maps of its surroundings, identify objects like houses, trees, and even other drones and birds, and then teach the AI reasoning skills to help it with tasks like identifying landing spots. In 2019, Daedalean is already a few steps beyond that — it is identifying landing zones and aborting landings when the status of the landing zone changes.
Chip in a few dollars a month to help support independent cleantech coverage that helps to accelerate the cleantech revolution!
Another thing Daedalean has taught its AI is to turn its visual mapping skills into landmarks it can use to navigate and find its way back home even without GPS. This also happens to be really cool to see. It may sound simple but it really isn’t. For more information about the technical challenges behind flight automation, make sure to read our piece “2019 Beginner’s Guide To Full Self Flying, & How It’s Not Like Full Self Driving.”
Daedalean’s latest demonstration, which CleanTechnica witnessed in person, is that its drone can scan the terrain and determine safe locations for emergency landings, and then constantly reassess whether those locations are still available, something the AI does throughout the entire flight. They demonstrate it by driving a van through the emergency landing zones and make the drone abort landing procedures as needed, all without human intervention. It’s a truly impressive demonstration, but they still have a long path ahead of them and so does the entire industry. Now that they have almost finished with vision, they will soon move on to giving the AI more experience in different urban and non-urban scenarios, and afterwards add more sensors.
Have a tip for CleanTechnica? Want to advertise? Want to suggest a guest for our CleanTech Talk podcast? Contact us here.
Latest CleanTechnica.TV Video
CleanTechnica uses affiliate links. See our policy here.
