Really impressive demos from Skydio in Redwood City. Quad copter startups have been hit or miss & it looks like the same pile of plastic scrap that every startup besides Apple produces, but what they showed here was exactly what I wanted to do for years, with ground vehicles. It looks like it could really navigate most of the mobs I have had to navigate manually. They showed it briefly flying in front of the athlete, despite not having prior information about where the athlete was going. The flight time is undoubtedly very short & the $2500 is almost 1 month of rent. As a $2500 consumer quad copter, it has no chance, but it would have a chance as a kit allowing developers to add autonomy to ground vehicles.
They're the 1st to consolidate a lot of the bits of research that were done in the vicon rooms for the last 10 years, into a complete product. They emphasized the demo of the fixed wing flying in a parking garage. It's the 1st time anyone has shown the NVidia TX1 doing anything besides raising its stock price.
Not sure how well it would translate to a ground vehicle, because a ground vehicle needs to stay on a certain path. Prior knowledge of roughly where the athlete was going & where the path was might allow it to do the job. Suspect GPS has not improved enough to stay on a path without some other input.
It wouldn't translate well to a Casey Neistat application. His interest is manely panoramas & very stable long shots. Skydiddy is aimed at very close, jerky closeups. He would complain about the motion being too jerky.
The mane development is not 4 but 13 cameras, not edge detection, chroma keying, sonar, or lidar, but depth mapping from parallax & optical flow alone. The cameras may be high speed HD. It takes an immense amount of power to combine 13 cameras into a very detailed 3D map. The ability to plan a route several trees into the future is another area which requires specialization in algorithm design.
You can't get a startup funded without mentioning AI somewhere, but they probably only use it for recognizing the subject to track rather than the obstacle avoidance. It apparently does a much better job than SURF, the hero algorithm of 5 years ago & now obsolete.