So my thesis is coming to its end. The new pipeline seems to be working. We get higher fps (~20) on bigger input images. In addition for control, camera capture and inference, we can now also stream the images from the drone's camera. So we made a video from the nano-drone's point of view. The weird thing on my forehead is the marker we use for the tracking system so we can compare the inference to the GT.

https://youtube.com/watch/...

That is so cool! I just wonder if you have ever been decked in the face before by that drone because of some weird bug?

@Mark-Zuckerberg So far bugs of various types made it crash into the safety net and the ceiling. One of my advisors is always present at these demos, with a controller that can send an "emergency landing" command. I got pretty good at diving and catching it before it hits the floor.
Anyway, it's a 27g nano-drone. No real decking danger.

I'm impressed. Is this for undergrad, masters or PhD?

For either of the first two options especially I have to say it's one of the most impressive practical projects I've seen. Good advisor + hard work?

@AlmondSauce Master. I did have a couple of years as engineer in the industry before I returned for M.Sc so it definitely helped. And yes, very supportive and technically-able advisors.

@NickyBones that is awesome, great work!

Do you have any future plans for it like creating a personal army of nano-drones or as a single product or something?

@bioDan That depends on whether I will get a phd grant :)

reminds me of the netflix show omniscient 😅

Wow, you know your way around CUDA. Amazing. CUDA always was something unexpectedly hard for me even though I’m fine with MMX

@uyouthe This project specifically does not include CUDA since we don't have GPU on the nano-drone. We don't even have FPU :)
I have worked with CUDA before, but the language evolved quite a lot recently.

@NickyBones I read a paper a while ago where they trained it not to fly, but to FAIL at crashing.

Apparently identifying the actions most likely to lead to a crash and rank ordering them is easier than identifying the actions that are most likely to sustain flight.

May have been a GAN. fuck if I can remember though.

@Wisecrack GAN seems like an overkill. But a relatively simple ResNet based NN that basically does obstacle avoidance is DroNet - they only calculate the probability of collision and the steering angle required to avoid the crash.

@NickyBones I always forget about ResNet. Such a remarkable design.

Also what ever happened to liquid state machines, echo nets, and reservoir computing in general? I haven't followed the state of research in a while, and these seemed to offer so much promise (theoretical universal function approximation for example).

@Wisecrack I have no idea! I am only a code hamster :)

@DevPidoroSan Most games barely break even, and that assumes you know what you are doing.

Get VERY familiar with marketing and networking.

@DevPidoroSan You'd be better served building in unity or unreal.

Javascripts good I guess if you want something that runs in browser, because thats a low barrier to entry, just dont expect much performance wise.

Also instead of relying on *one* youtuber to pick it up, start making lists of youtubers who play the sort of game you're making, usually anyone above 100k subs is a long shot but include them too. Some people say participate on the channels or discord/twitter/etc on a semi-regulary or daily basis so you're not just popping in to ask for free advertising, but that seems disingenuous to me. What I would do is, the channels I already like, I'd start participating with more, not a lot of time per day, just maybe 5-10 minutes, an hour on the weekends in their discord or whatever, get to know whos who, say hi, etc.

Most of all if you're new to development, the best thing you can do, despite *all* advice to the contrary, is make the game YOU would want to play.