3D Robotics

Using AI to design a drone frame


This render from Fusion 360 features the Hooligan 1000, with an additional thruster and rotors set at 30-degree pitch for optimal thrust. The final design may change based on the results of Autodesk Dreamcatcher. / Credit: Eli D'Elia, Autodesk

From Core77:

Someday in the not-too-distant future, a drone designed by a computer will be flown by a computer. That's Eli D'Elia's dream. 

It's a dream he's working to make a reality by partnering with Autodesk to design the Hooligan 1000, among the first entrants in the newly emerging 1,000 mm drone racing class. But racing is just the beginning for D'Elia, a product designer, roboticist, and professional drone operator who has helped to pioneer drone racing as we know it. For him, it's more important that the Hooligan be durable, practical, and easily adapted for use in industry and agriculture, possibly search and rescue, and eventually, mapping and exploration. He envisions a future where drones move not only through the air, but through water and outer space as well. "We're really just in the Kitty Hawk stage of drone development," says D'Elia. "Competition through sports is a great way to test out ideas and get rid of the bad ones."


Letting the Computer Do the Work

"AI-grown, AI-flown," is how D'Elia summarizes his goal. To that end, D'Elia and his partners at Autodesk, Taylor Stein and Daniele Grandi, will use the Autodesk Dreamcatcher generative design system to design the drone's chassis. With generative design, the designer doesn't come up with the actual design, but instead inputs a set of requirements and lets the computer go to work. While still in its infancy, it has already been used to create an experimental automobile chassis and components for commercial aircraft.


After the design is finalized, D'Elia will start working on the AI control system that will steer the craft using 3D photogrammetry and the NVIDIA Jetson TX1 processor. He expects the AI to be able to learn a given racecourse after several test flights, then fly it both with and without human control. But he also wants it to have situational awareness, the ability to identify objects in the air like birds and other drones, and respond appropriately. "We're putting an additional thruster in there to provide a turbo boost at the end of the race," D'Elia says. "I'd like the drone to be able to tell if there's another drone close to it so it knows when to kick that thruster in for some extra speed."

Drones Far and Near

What D'Elia wants to do and what the Federal Aviation Administration (FAA) will permit are two different things, of course. According to current FAA regulations, all drones in the U.S. must have a pilot at the controls and be within the pilot's line of sight. And it may be years before autonomous drones can be flown in civilian airspace simply due to the danger of a drone hurting people, damaging property, or interfering with air flight. "They'll need to do things like employ redundant flight control systems so that, if a propeller goes out, it won't fall out of the sky," says D'Elia.

But there's no question that many people and companies see a future for drone autonomy, both near and far. Amazon is testing delivery by drone in the UK, where more lax regulations permit drones to be operated beyond the pilot's line of sight. Meanwhile, the startup, Zipline, is already using autonomous winged drones to deliver emergency medical supplies to distant communities in rural Rwanda. And companies in the U.S. are petitioning the FAA for waivers to the line-of-sight rule. The first such waiver was granted in August to PrecisionHawk, an aerial data analysis company, after they provided a year's worth of drone flight safety data.

For D'Elia, all of this is interesting, but the truly exciting possibilities lie further afield. He notes that there's already an XPRIZE competition underway to design a drone that can map the ocean floor by itself. D'Elia thinks it's only a matter of time before similar efforts begin to explore outer space. 

In the Field and in the Lab

When he's not working on the Hooligan 1000 designs, you'll often find D'Elia flying drones in the vineyards of Northern California. The company he started, Eagle Eye Metrics, deploys drones to help farmers, generally vintners, monitor their crops and map their fields. Using a NDVI camera, his drones capture infrared frequencies that can reveal information about soil quality, pest infestation, and overall plant health. On foot, this kind of necessary recon can take farmers a week. D'Elia and his drones can do it in less than an hour. 

531403_4_57167_Dih6YzsPq.jpgWhen he's not designing drones, Eli D'Elia performs drone-based aerial mapping for farmers in Northern California through his company, Eagle Eye Metrics. / Credit: Eli D'Elia

D'Elia has always had an interest in both the sport and the practical use of robotics. Long before BattleBots hit the airwaves, he was part of an amateur robotics community in the San Francisco Bay Area in the '90s. They would build their robots during the week, then put them into the arena to do battle on Saturday night. "The whole idea was to one-up whatever the other guy's robot did. If he clobbered you, you hit him with a projectile. If he fired at you, you hit him back with a flamethrower," D'Elia says.

When drones hit the hobbyist market in the mid-2000s, D'Elia was immediately interested. However, he and his frequent collaborator, Marque Cornblatt, saw one big problem: drones were too fragile. "You could spend two weeks and $1000 building a drone, but you can make one mistake and crash it in ten seconds and you're back to square one," D'Elia says.

Under the auspices of their own organization, the Aerial Sports League, D'Elia and Cornblatt designed the Hiro, a drone with a monocoque frame made from a lightweight, super-strong polycarbonate used in aerospace and military applications. The design was fireproof, waterproof, and sturdy enough to withstand direct hits from a baseball bat and even a shotgun blast. It was among the first successful drone projects on Kickstarter. 

D'Elia and his Autodesk collaborators are currently finalizing design parameters for the Hooligan using Autodesk® Fusion 360™ software, specifying requirements for mounting connections and access ports. When completed, they'll feed the requirements into Dreamcatcher, which will take several weeks to generate designs for the chassis. They'll then print that design in polycarbonate, assemble the components, and take it out for a test flight. Then they'll begin working on the pilot AI with help from NVIDIA. 

D'Elia remains upbeat about the near future not only for his project, but for drones overall. "Sometimes you hear people say that all this amazing stuff is going to happen 'in the future,' but you have to remember—that's like three to five years from now," he says. "Think how far self-driving cars have come in five years. This stuff is coming and it's coming fast." 

Eli D'Elia and Taylor Stein will be speaking about designing racing drones using Autodesk Fusion 360 software and Flow Design wind-tunnel simulator as part of Autodesk University 2016 in Las Vegas, November 15-17. Learn more and register today.

E-mail me when people leave their comments –

You need to be a member of diydrones to add comments!

Join diydrones


  • @JB

    Hi JB

    Your MINI Talon VTOL is a very interesting project. I would be very grateful if you can share more details of your configuration.

  • Ok Cool. I sent you a PM with my email last night. (AU time)

  • Yep, I have an Ultimaker.

  • No probs Rob. Will try to get it written up within the next week or so. I'm currently busy quading some Disco's for Parrot, which I'm hoping to add some 18650GA battery as well to double the range. I assume you have access to a 3D printer? I'll PM you my email so I can send you the STL's.

  • Actually, I would really appreciate the build details (and soon) as I need to build a small QP system for some development work.

  • Rob,

    I don't have any videos in wind with the current airframe we used at the OBC in September this year. This is still a work in progress, but I do have a video of our maiden landing in over 30kmh wind, without windvaning and forward motor functions (as this was not part of Arduplate 3.5 in Feb16 when we took the video for D2) you can see here: https://youtu.be/5n6W8VV-bPk?t=248

    Note in this video it does not have forward tilted quad motors, hence the wing is creating downward lift with it's negative AoA. The current Talon QP version has 7degree forward tilt that means the wing always produces lift in hover with wind, when the forward motor is used for position hold. 

    The quad arm propulsion only weighs 100g each (200g total QP drivtrain), including 3D printed motor mounts, CF bar, ESC and 2206 motors with 6x3.5 props. Hover is at 40A and forward cruise with the quad attached is under 5A at 20m/s running on a 4S 10Ah LiHV. Quad thrust is 4.4kg with a 2.5kg MTOW. There is only a tilt gimbal for the pi camera, as that is all we required for the OBC (ie this is a long range survey airframe). The forward propulsion is optimised for speed/range, rather than static thrust, as the QP launch means it has a "unlimited runway" in the sky to takeoff and land with. Because of the QP we can also put more batteries in than a normal mini Talon and only have to run with a bit more airspeed to compensate for the extra mass and much higher wing loading (which also helps in wind).

    We're planning on providing the build instructions, including the 3D printed part files etc, for the mini Talon QP conversion, so that others can use the platform as a electric long range (100km) SAR VTOL airframe. 

  • JB,

    You would need to be able to yaw all the way around for many applications.  Inspecting a tower, for example.  At least as long as the camera is mounted in any kind of streamlined fashion.  Unless you are planning on mounting it on a 360 yaw gimbal mounted below the fuselage, which would be an interesting arrangement for an airplane...

    And when you quote these power figures, that is for how much payload?  Most applications I have seen of quadplanes have very limited payload.  This is capable of lifting 4kg.  Actually, could do 6kg if you remove one battery and "only" fly for 15-20 minutes.  And this is out of an aircraft about the same size as a Skywalker.  Do you have a Skywalker VTOL that can lift a real camera on a gimbal?

    Quadplanes, while interesting for some applications, are very payload limited, as much of the payload capacity of a given airframe is taken up by the inefficient lift system.

    Further, as John mentions, do you have any real video of a quadplane hovering in 50 km/h wind?  How about on-board video from a gimbal in such conditions?  I'm just getting started.  Waiting for 80 km/h wind.  Any video of a landing in these conditions?

    You know, I flew planes for years.  I know how they are in wind.

  • Developer

    At the relatively low altitudes where we operate, wind will be gusty and turbulent more often than not.

  • Hmm Rob.

    Why would you want to yaw the aircraft when using a camera gimbal for inspections? I would have thought holding position would suffice, provided the camera and gimbal have clear view of the inspection task? 

    Although the video and Heli is impressive (I'm a fan of your work) our quadplane would be hovering there on about 4A (on 4S) in 50kmh wind using wind feathering and forward motor for position hold. 

    Quadplane performance improves with wind, to the point that over 9-11m/s it hovers without quad motor lift, which is lower than the wing stall speed of the aircraft. The reason for this is that the forward motor is maintaining lift and position, whilst the quad motors are only being used for attitude control (along with the wing control surfaces), resulting in low current draw in wind. Because of the quad assist the wing is unable to tip stall, and high AoA pitch attitude improves lift further as the quad assist dynamically prohibits stall as required. Gusty, variable direction wind is more challenging (as quadplane yaw control is fairly low when using small 6x3.5 props on a 2.5kg frame) but windvaning still works nontheless at the cost of some momentary stablity only when the wind changes direction. 

    In fact at 9m/s it becomes difficult to desend, and is something we still need to address in Arduplane code (ie to use quad pitch control to descend with forward motor on) You can try this in SITL by running the quadplane model and setting the wind to 9-10m/s, with windvaning and forward motor assist on. Flying in wind over about 13m/s results in a "climbout flyaway" until the battery is depleted, unless manually recovered.


  • Quadplane can't fly properly.  ;)

This reply was deleted.