Lessons for deep-tech startups from Skydio, the autonomous drone company

Lim Zhan Wei
Feb 23, 2018 · 6 min read

My hypotheses on how Skydio delivered a deep-tech consumer product in an once over-hyped and now extremely crowded consumer drone space.

Skydio recently announced R1, a fully autonomous consumer drone that can track a person. While person tracking in drone is not new, R1 is said to be way above its competition. You can see in this video that R1 did not lose track even when the reviewer ran in opposite direction and occluded by trees for a few brief moments while trees in its path. R1 certainly uses cutting-edge technology. It has 13 cameras for simultaneous localisation and mapping (SLAM), has Nvidia TX1 onboard and uses deep neural network to identify the person to track (so that it won’t track the wrong person). What’s most striking to me was that the technical building blocks for this engineering feat have been becoming more accessible in recent years. If the building blocks are there, how did a startup like Skydio manage to beat big company like DJI in bringing an advanced product to market? I believe deep-tech startups can learn plenty from Skydio.

Company vision

Skydio was founded in 2014 by Abraham Bachrach and Adam Bry. Skydio set out to be a fully autonomous drone company. It was a time when quadcopter is in the highest point of the hype cycle. Many drone startups have failed since (Lily robotics, 3D robotics, etc.). DJI came out as the dominant player from that cycle, winning the competition on price, quality, features for almost all product categories.

Many drone startups also claimed autonomy as their competitive advantage but they often lacks the conviction to deliver the best autonomous experience. For background, there are many different levels of autonomy a drone supplier can bring to market. Flying in a straight line within radio range is considered autonomous and the simplest to implement. You can add a few sonars to prevent the drone from crashing into the wall. You can also add a camera to help stabilization, and then add another for person tracking, and so on. The more steps you take towards autonomy, the more sensors and computation it needs, resulting in greater power draw, shorter flight time and lower payload. And of course, the drone becomes much more expensive due to the cost of sensors and sophisticated onboard computers.

The most valuable customers to drone companies are hobbyists or professional photographers who knows how to fly a drone already. The rest who can’t fly well probably kept their drone in the closet after a few flights and never buy another drone. The most valuable customers care most about flight time, payload (whether it can carry a gimbal, professional DSLR,etc.), and price. Increasing autonomy, something valuable customers don’t need, has a direct impact on the metrics what valuable customers care about.

Furthermore, autonomy is hard to build. Though I claimed that the blocking blocks are available (more in the next section), it is still hard to put them together in a hardware product under severe power and weight constraints. With a tight timeline, autonomy often becomes an afterthought for drone startups. Most just slap on the minimal level of autonomy it can get away with, just enough to look good in demo videos and not be called a fraud.

Skydio’s conviction around autonomy is clear in its latest product. To carry 13 global shutter cameras and Nvidia TX1 board, and to run neural network for person tracking and movement predicting and SLAM algorithm for obstacle avoidance, means R1 sacrifices on nearly all other aspects of a quadcopter.

Lessons for deep-tech startups: Like autonomy in drones, many deep-tech startups need long-term commitment to execute its vision. This means sacrifices in product features when compared to traditional product and long term investment in R&D to tackle hard engineering challenges. For deep-tech startups operating in an over-hyped space, there are plenty of opportunities where it looked like profitability is just around the corner. It is classic “the grass is greener on the other side”. When faced with a hard engineer obstacle, it can be tempting to take the seemingly easy way out. However, the easy way out in a hyped-up space is often a path to failure as it leads you right into middle of the fiercest competition.

For example, Skydio could have dial R1's autonomy down a few notches to match price point of a DJI hobbyist level drone and still have slightly better autonomy than DJI’s drone. The drone will look more like DJI’s, and it eliminates huge R&D cost of developing full autonomy! But it will end up competing right in DJI’s territory. DJI could easily bring the down price of its competing drone (happened before) to kill the startup, and the slight improvement in autonomy will not save it.

Riding on waves of innovations

What’s more impressive is to consider the technology landscape back when Skydio was founded. Vision-based SLAM just started to look promising, but LIDAR is still the go-to sensor for SLAM (it still is today). Deep learning was at its very early days. There is no such thing as an embedded computer that can run deep neural networks like Nvidia TX1. Which means person tracking on a drone in 2014 is never going to use a deep neural network. It would be impossible to build Skydio’s R1 in 2014.

Fast-forward to 2018, most of the technology building blocks for R1 are out in the wild. There are open source codes for vision SLAM on drone and for person tracking (try Googling it). The various open source implementations out may not be the best or reliable, but the methods for achieving R1’s features set had been figured out over recent years. It is much easier to build an autonomous drone today than in 2014.

While I do not have insight into how design decisions have evolved, it is likely that Skydio bet on computer vision at the very beginning and the wave of innovations in computer vision carried it through. This doesn’t mean Skydio’s engineers have been sword-fighting in office until the right technology pops up. I believe they place long-term bets on vision and have been working hard to close the gaps between state-of-the-art of the day and performance requirements of its product. Their work gets easier when state-of-the-art improves. No engineering team can pick up a new technology at moments notice, especially it has to be tightly integrated into a hardware. They need to be experts in those area of technology to work with the state-of-the-art, and the yet-to-be invented technology has to be on its roadmap.

Back in 2014, they could have opted for using a tracker device that the tracking subject carry to do person tracking (Lily robotics was based that). As it plays out, computer vision experienced rapid progress with deep neural networks, whereas RF technology remains largely the same over the same period.

Lessons for deep-tech startups: One popular playbook for deep-tech startups is to possess or develop a highly specialized and fundamental (hence the word “deep”) piece of breakthrough technology. This piece of technology then becomes the startup primary competitive edge and which it relies on to sail through its competition. In this narrative, deep-tech startups are monopolies over fundamental technology. This plays out nicely for startups like Google with its PageRank algorithm and more recently Velodyne with its LIDAR technology, and Magic Leap (if its headset works as advertised).

I don’t think Skydio has any monopoly over fundamental technologies in its drone (person tracking, vision SLAM, flight control, etc.). Its defensibility comes from the complexity involves in bringing multiple piece of cutting edge technology on a hardware platform.

In rapidly advancing fields, such as AI, robotics, there are more opportunities in riding on waves of innovation than seeking a monopoly over certain piece of specialized and fundamental technology. This means that deep-tech startups in these should not compete with big companies, research institutes, and university over breakthrough in fundamental technology (for e.g., build the next generation deep learning models,etc.). Deep-tech startups’ role is to deliver breakthrough product which certainly exploit the best technology of the day. However, to ride on the waves of innovations, the team need research level expertise (crudely defined as able to publish in top tier conference if they want) to understand and exploit cutting-edge technology. For instance, while many open source code are available for SLAM, without an intimate understanding of the underlying algorithm, it impossible to make improvement in any direction, say to lower power consumption, etc. If the product called for, they need to do research to close the gap between state-of-the-art and product requirements.

Deep-tech startups should actively bet on what is going to happen in their field in the future. Being fixated to what technology can offer today means risk missing the waves of tomorrow and become irrelevant.

Without first hand information about Skydio, I can only guess the various circumstances based on my understanding of SLAM and person tracking technology and public accounts of Skydio. (So please correct me if you have more information or think I got it wrong, I’ll update this post accordingly.)

Lim Zhan Wei

Written by

trying to write something worth reading while I'm not doing something worth writing, and vice versa.