This looks great! I really suck at making cinematic videos with my drone and I would welcome the possibility of just instructing it in natural language.

Really cool engineering feat!

I think you're making your route to market overly complicated by calling it autonomous. Just call it what it is – a drone with a AI navigation system. As long as a person commands it and monitors it (and thus _is responsible_), there's no reason to call it autonomous.

First video has 131 views – I hope this blows up.

If you're targeting first responders, you could approach private ones first (festival, big event, fairs, etc..). They have basically the same needs as the public ones, but their operation theatre is sometimes/often a private one (private property) with a lot less regulatory overhead.

Obligatory IANAL here, but (not US) event organiser.

In the 60s we could do a small amount of basic numerical compute with tubes, relays and crappy germanium transistors.

From day one, there were use cases for computers, as long as they provided tons of value with very little compute. It took until the late seventies for computers to be cheap and useful enough lower-value small-scale use cases.

Particularly banking could extract a lot of value from, say a 1 Hz CPU, because human computers more of a hassle to deal with than 9000 inefficient tubes or relays In the 60s, using computers to decode h256 video to watch some cooking instructions would be absurd.

Today, we extract value from drones, at a certain cost. The cost is determined by how much of a pain in the ass it is to use. For good PMF you need to get the value to pain ratio high enough.

These are the costs/pain points that hold back further adoption, in my personal experience:

Flight planning:

- Check adjacent land use (recreational, harbor, airport) for regulatory reasons.

- Check NOTAMS and airspace restrictions.

- Obtain airspace clearance (requires a phone call currently I believe)

- Check wheather, esp. wind speed.

- Ensure ground area is controlled and all people are informed.

During operation, continuous observation from operator required in order to detect the following potential issues:

- Battery level critical during RTH (time between low and critically low battery voltage can only be estimated and is affected by battery wear).

- Compass failure. State estimation algos will reject bad compass readings and fall back to GNSS for yaw estimation, so this can be avoided with good code)

- GNSS failure (can always happen due to tall obstacles)

The ground area needs to have no uninvolved persons (unless drone has safe abort system such as a parachute), because the following could cause an unplanned meeting with the ground:

- Battery failure

- Airframe failure

- Collision

- Thrust system failure (prop, motor, inverter)

- Flight controller failure (easy to lock up the MCU with some bad code)

If the above was not a reality, we would see more drones, esp with full autonomy. There are a lot of possible failures that are hard (but not impossible) to automatically detect and mitigate. Effectively constant monitoring by an operator is needed (this can be done remotely from a control room).

I see that it uses cloud compute. Would a local multi-model be impossible or just too limited?

Maybe reach out to these guys (https://www.breakerindustries.com/)

Looks like you are walking down the same path in different places.

You could take Uber's approach. Ignore regulations. Focus on delivering such massive value that the regulations are changed to suit you.

Cool demo! There are immediate use cases in the farming space, though competition is strong. Happy to chat.

lol I just made a ask hn post asking why we don’t see more stuff in the drone space.

Are you a software engineer? What’s your background? Cool demo