I interviewed Michael Dewhirst from Evolve Dynamics. He’s one of the people who is working full time on the drone revolution. I thought it would be interesting to hear from an “insider” on the challenges for the technology – as well as getting the details on just how far he thinks this transformation can go. This interview is all about the thorny problem of control: how we manage thousands of drones, all criss-crossing the sky at the same time.
AL: How did you get into commercial unmanned aerial vehicles (UAV) – also known as drones?
MD: It all started with a hobby drone a few years ago. After that, my other hobby of rock climbing got me into professional film making. I found that many issues were unsolved: how to fly a heavy camera in difficult conditions, often without a GPS signal to help keep the drone steady, as well as flying for a while – eg, more than 15 minutes.
After that I founded Evolve Dynamics to work on the many problems and – as I see them – opportunities within the space.
Entrepreneurship is what I’ve been doing for the past 20 or so years, and I’d previously worked for BAE Systems.
So far we’ve built a great team at ED with aerodynamics engineers that have 25+ years’ experience, mechanical engineers, circuit board designers and high-end software developers.
We’ve built a commercial, mid-size UAV which flies for 30-45 minutes with a 10kg payload and it’s very quick to fold up and fold out for transport – as well as the battery taking less than ten seconds to swap out.
AL: What future lies ahead of us in the next decade or so, as far as drones are concerned?
MD: I see a future much like the one in Back to the Future Part II or Futurama, with thousands of flying personal transports – clean, electric, mostly solar powered. I see unmanned cargo transport – also electric and solar powered and the end of diesel trucks. We are not there yet and we have a while to go yet to get there, but there are inroads being made in many areas. Electric motors are becoming more efficient every year, and my company has been working with a great developer of such motors – KDE Direct.
There is also much more work that can be done with propellers and aerodynamics, such as ducted fan engines. Think commercial airline motors with the “pipe” around a propeller – this greatly increases motor efficiency and means an aerial aircraft can fly for much longer. Aurora Flight Sciences recently released their scale-model LightningStrike aircraft that uses a cluster of 24 ducted motors and which won $89 million USD from Darpa to take the concept to production.
Battery technology can also be pushed a lot further. New technology is being discovered every day, such as graphene and organic polymer saline redox-flow batteries which have emerged in the last year. Another major area of work is autopilot systems. There are a number on the market, but a lot more development opportunity remains. One of our competitors, 3DR, have a great product with their Pixhawk autopilot module – they even support some rudimentary visual orientation and sonar based collision avoidance.
DJI also have several autopilot modules with the A2 and WooKong – as well as add-on modules such as Guidance and Manifold that allow developers to work on visual guidance systems. Autopilot systems are the most critical component in the puzzle as they are literally the only way the whole “flying cars everywhere” future can exist. Humans currently drive – and crash – cars, and these go along a single line most of the time anyway. Imagine what would happen if we get people to pilot a flying car, with thousands of other cars in the sky.
The only solution is a sophisticated, safe autopilot that would pilot such “cars”. A networked system is not really viable due to possible connection issues, as well as security. The autopilot would need to be self-contained and capable of “seeing” in 360 degrees and three-axis around itself, and we are working on such a system at Evolve Dynamics under the code name of Exa Iris.
There are already some early prototype personal transport systems such as the EHANG 184 which can carry one passenger for 20 minutes. The company plans to pilot the aircraft via a central “air traffic control” hub which has not been built yet. Joby Aviation has also been working on a two-seater, hybrid vertical take-off winged craft – long range personal aerial transport that is electrically powered and can fly at speeds of 200mph and has a range of 200 miles. Finally there is the Volocopter, which has 18 motors and can carry two people for 20 minutes but has manual controls. So we can see that all the components needed for the future I painted are there, and just need further work to crystallise into a complete solution.
AL: What other challenges do autopilot systems need to solve, especially in your future with thousands – perhaps millions – of flying aircraft?
MD: Coordination between many autonomous aircraft is a hard one to solve – when you have many objects moving through three dimensions, each with their own agenda, you have to make a lot of calculations in real time. Hefty computing power is needed to power machine learning algorithms and artificial intelligence, as well as the ability to process a lot of data. The problem with computing power is weight and power consumption. The more you want to be able to process, the more the computer weighs and the more power it consumes. In the air this is a problem because weight is a major issue, especially when dealing with small aircraft.
AL: Aren’t computers getting smaller and lighter now?
MD: Indeed they are but not small enough for UAV and drones. There are, however, also specialist SBCs (single board computers) out there. They’re much smaller than a typical computer motherboard. These boards are more compact, drain much less power and allow very customised applications development. Beyond the reasonably powerful “standard” SBCs with a CPU (central processing unit) such as Raspberry Pi, Banana Pi, Udoo Quad, Parallella and others, Nvidia has its Jetson board that boasts a very powerful GPU (graphics processing unit) perfect for video processing and is at the heart of DJI’s Manifold.
Movidius has also just released an innovative “clean slate” architecture system called Fathom which has been developed specifically for artificial intelligence and video processing applications – small, compact, powerful and also needing under 1W of power – which is pretty unprecedented in the SBC world with boards capable of this sort of power. The Fathom was at the heart of the DJI Phantom 4 drone, which is the first consumer drone to have front-facing stereo vision cameras.
AL: What possible issues would autopilot systems need to deal with?
MD: In dense formations, if the coordinating air traffic control system went down, the whole swarm of airborne aircraft could go into an escalating “pile up” type multiple collision. The air traffic system would be vulnerable to malicious hacking attacks and viruses, as any “connected” system always is by its nature and purpose. A remote operated US military stealth drone went down a few years ago and it was speculated that it was deliberately downed by government hackers. One possible scenario is if the autopilot and its backups failed – that’s highly improbable, but has catastrophic consequences. Even if it was not possible to save the failed single aircraft, others would need to respond quickly.
The best solution is a triple redundant architecture. If the autopilot units were capable of operating on their own, as well as in unison – not just by following instructions from an “air traffic controller” hub – they would talk to each other and coordinate movement without input from an “external observer”. If for whatever reason this was also not possible, the third method of operation would be fully autonomous flight control – where the aircraft would monitor visually and through other means the position of other aircraft around it and adjust trajectory to avoid collision without any communication. This is obviously the most complex.
AL: You mainly talk about fully autonomous flight. What about situations where the person inside the aircraft would want to control it in real time?
MD: You would need a system that was capable of allowing control within a safe corridor. A bit like remote controlling a train. You can go back and forward, and turn at junctions, but not come off the track. The autopilot would allow the user to direct the aircraft as desired – as long as it was within the “safe corridor”.
In open air and terrain the only limitation would be obstacles such as ground, trees, mountains, buildings and so on. When in danger of colliding with any of these, the aircraft would simply stop and hover, or adjust course. In a busy, congested corridor, the vehicle would most likely allow very little input – due to other vehicles and obstacles around. This would be like motorway driving; you can only come off at an exit.
AL: Are you also working on coordinated flight?
MD: Not yet, this is a few years away yet.
AL: Speaking of busy, congested air corridors, how do you think this will be managed in large cities?
MD: One solution would be to use ground beacons which would constantly transmit local traffic corridor configurations – kind of like digital lane markers. Without such predefined flight paths it would be impossible to coordinate, or expect aircraft and their autopilot systems to self-coordinate. In less congested areas, in the absence of guidance beacons, it would also be possible to pre-program artificial intelligence algorithms on board the autopilot to default to certain behaviours – a bit like the Highway Code.
AL: What is the current main limiting factor for consumer and commercial UAV and drone usage?
MD: Most drones still require skilled, trained pilots to operate them safely and correctly for good results. There still is a big gap between the automation that the autopilot provides and the manual skills required to safely fly.
AL: What are the main problems that don’t yet allow us to have highly automated flight?
MD: UAVs and drones currently on the market are limited by their ability to get a good GPS satellite signal – such as indoors, or on streets with taller buildings. There are some emerging technologies, such as the DJI Phantom 4 and a startup called Augmented Pixels, which use machine vision cameras for obstacle detection and collision avoidance – but both have limited range. Most current solutions are only front and down looking and are oblivious to any other problems – like if it has to “reverse” or ascend and goes into a roof, ceiling or even a person standing behind it!
AL: What is the problem with range?
MD: If the UAV flies higher than 15 metres above the ground or further than 15 metres from the nearest front-facing surface, for example a rock face, and there is no GPS signal (because you’re in a canyon or a large hangar), the UAV will not be able to understand where it is, would no longer be able to hover steady, and it could crash.
AL: Is Evolve Dynamics working on solutions to these problems?
MD: We have indeed been working on the next generation of autopilots at Evolve Dynamics, a system called Exa Iris, which can operate without any GPS signal, as high as 100 metres above ground or from any surface and still hold a steady hover regardless of wind.
It is also designed to detect stationary and moving objects and facilitates collision avoidance in all directions, and has no “blind spots”. Exa Iris also generates a 3D model of the environment it is in for immediate or later use. The model could then be used to pilot the UAV to a desired area without manual controls – through simple “point and click”.