Four months ago we launched the Skydio R1 – a fully self-flying drone that we declared the most advanced autonomous device of any kind available to consumers. The customer and press feedback have supported that claim, and judging by the number of (misguided) Skynet references, so have internet commenters. It succeeds in bringing robotics and AI research to market in a powerful way, conveying the unmistakable potential of the underlying technology.
As a startup creating a first-generation product, we strove to deliver a breakthrough on one axis rather than immediately trying to cover all use cases – making the leap across the uncanny valley of semi-autonomous features to deliver a new type of experience, one where the tool becomes trustworthy enough to make its own decisions. We did this with ten world-class people working on the core autonomy system. In the process, we’ve learned what it takes to develop hard tech and how to accelerate progress in robotics research (interesting external perspectives here and here).
To ship an autonomous consumer product demands a level of robustness in geometric and semantic understanding only recently made feasible with advances in robotics algorithms, embedded processors, and machine learning. It means real-time omnidirectional perception and action. It means dealing with things like smudged or broken lenses, sun glare, reflections, wind, snow, and dinosaurs. It means handling uncertainty on every axis – the R1 may not understand what’s happening the next time you take it to a house of mirrors, but it’s smart enough to act conservatively. In the end, the internal complexity must abstract away to provide a seamless, simple user experience.
Our strategy to achieve this is simple – a small team of extremely dedicated and talented people with a foundation in math and the ability to learn, an appetite for long-term risks, and the willingness to rip something up and start over when trapped in a local minima. Our culture is built around reducing the cycle time from idea to flight, which is supported by our vehicle fleet, simulations, and large-scale log data. There’s a spectrum of consciousness in the blend of research and product that reminds me of this fantastic post from Tim Urban – at the base of the engineering staircase are bug fixes, then parameter tweaks, then adapting literature, and finally exploring novel ideas. Success depends on how much time smart engineers that understand the big picture can spend at the top of that staircase.
Over the next few years, robotics systems will make significant leaps towards total scene understanding at superhuman levels. Drones are the perfect platform to conduct this research. Self-driving cars offer huge potential, but so much of the current development effort goes into leveraging prior maps, exactly following the rules of the road, and handling the fifth sigma of edge cases from human drivers and pedestrians. Drones are small, relatively safe, and the navigation problem is very general. To film a soccer game or scan a bridge effectively requires creativity and understanding the relationship between people and objects at the highest levels. To barrel down a forest behind a biker requires tight prediction of the future and mastery of dynamics. At Skydio we’re working on the next generation of algorithms for achieving total scene understanding and pushing them into real products because we believe drones will be a ubiquitous computing platform for the physical world.
We are lucky enough to live in a time of breakneck progress in robotics. It’s intimidating to keep up with the torrent of literature, even if it’s your job and you scan arXiv daily. As we head to CVPR this week and take in the nearly 1000 accepted papers, we’re excited about the trend of combining learned feature representations with geometric structure. We’re excited about more efficient ways to leverage huge amounts of semi-supervised data. We’re excited to meet more people focusing on real-time applications and real-world robustness. Most of all, we’re excited that we can finally give public demos, so swing by booth 137 if you’d like a glimpse into our self-flying future.
Hayk Martiros is the head of the autonomy team at Skydio, which brings together expertise in state estimation, dense mapping, subject tracking, planning, and control to deliver breakthrough autonomous flight on a shipping product.