Hacker News Reader

Comment by alsodumb | Hacker News Reader

That's true, but I'd call the smartest robot available to consumers that is relatively affordable.

Their tech blows my mind. They do all the navigation and motion planning based on just RGB cameras, running multiple neural networks in parallel on board the drone. All the depth estimation, including small obstacles likes leaves and power cables, they are doing it all onboard. I don't think anyone in industry come close to that.

I so regret missing interning with them when they were starting to grow in 2018 - they came to my University for recruiting and I was just starting my PhD and decided to not look for internships.

View on HN

> I don't think anyone in industry come close to that.

Check NVidia. I believe Jetson Nano can do stereo matching from multiple camera pairs in real time. From here one step to 3d points cloud. New Orin will have like 10x performance. Coming in January. So, it's within the reach. I'm thinking, have big stupid drones collecting the dust...

Skydio is too small to design their own chipset - they use Nvidia GPU onboard.

My comment was about people using neural network models for everything needed to fly a drone. I've seen models fail badly in ground based robots. They are flying a drone that's running so many models to be able to track, detect, localize, plan and fly. It's insane.

Jetson Nano can do stereo matching, and perhaps one can manage to even run multiple neural networks at the same time in inference mode. Getting it into production on real world systems is a whole different story. I can't imagine how Skydio manages to produce such robust and consistent outputs. I mean it kinda makes sense why Skydio is the first and the only company so far who's able to do it effectively - Adam's (CEO) been working on it for like more than 10 years now and his PhD thesis at MIT is awesome!

It's getting easier and easier to make such a drone. I mean it's within the reach of a single developer with $1000 budget. When Skydio came out it was some sort of break through (if we forget DJI and Autel), but now it's not that impressive, sort of normal. There are many models and libraries. NVidia has the whole zoo of them optimized for small devices. For free. I'm pretty sure even my level is enough to make a flying obstacles avoiding drone. I would use of the shelf hardware, of course. Objects tracking, waypoints isn't a problem. BTW, this would make a nice open source project. The problem is, in wrong hands it can do a lot of damage.

We don't have to forger DJI and Autel - Skydio beat them in autonomy. Period. You seem to lack the understanding of fundamental difference between how a DJI drone flies around and how a Skydio flies around.

I work on drones for research. I do RL. No one in academia can build something so robust. You seem to think that flying drones autonomously in the real world is same as running an RL simulation. Good luck trying to build it for a 1000$ lol. Sure, go ahead and try your open source project. Good luck to you!

I strongly suggest you to do some reading before talking about things you clearly don't know anything about.

My Autel EVO II flies impressively well tracking me in dense forest. May be not as good, I don't have Skydio, so can't say. All I'm saying is that gap between advanced hobbyist with off the shelf components and one of the best on the market is just a few years. Cheap mass production is a completely different story. And not surprisingly there are many startups with new ideas. I may join one one day. (my stereo cameras are coming, Jetson is waiting)

Sounds like you have a good startup opportunity here. For $1000 you could single-handedly replicate a multi billion dollar business. Why don’t you get started already?

It's not a replication. Only part of the software. Flying obstacles avoiding drone have been in labs for years. As for hardware nobody can beat China where almost all components are made. I can put them together in custom drone, many do it. You can even by a kit on ebay. It will have controller which takes care of balance. It can handle many physical configurations and hover drone in place. All that is needed is external 'decision making device' which will replace the human pilot with remote. Obvious choice is Jetson Nano. There is nothing new here today, that's what I'm saying. Just put things together. To make it work comparably to commercial drone will take significant efforts. Not that much to beat 5-10 years old products. Not sure they even existed 10 years ago.

The funny thing, I may try it. I'm more interested in 360 degree 3d scanner. But having that it's easy to put it on drone. I have a bunch of them, so it will be almost free.

Please let me know how it goes!

Not surprisingly the Skydio 2 robot is built using an NVidia chip.

https://support.skydio.com/hc/en-us/articles/5338292379803-C...

Reply on news.ycombinator.com