Original post
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)#291
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)#291
TACIXAT@TACIXAT
idk you can train drones to fly on a 3090 in a couple of hours, this is a 3x64 nn
4:52 PM · Jun 8, 2026 · 233.9K Views
Systems engineer Yacine claims drone flight policies can be trained in two minutes from pixel inputs on consumer hardware
The claim challenges a multi-hour 3D simulation baseline.
411.6K35677220.4K
Sentiment
Positive users celebrate efficient pixel-based drone policy training on a single 3090 GPU as a refreshing accessible breakthrough, while negative users dismiss it as impractical for real aerospace use or overly tied to LLM hype.
Pos
50.0%
Neg
50.0%
12 comments with sentiment.
Cluster Engagement
Posts from X
Most Activity
Most Activity
VIEWS13.5KBOOKMARKS60LIKES163REPLIES6
kache@yacineMTB
I can train the same policy in 2 minutes, policies that are trained off of pixels
TACIXAT@TACIXAT
idk you can train drones to fly on a 3090 in a couple of hours, this is a 3x64 nn
7hViews 13.5KLikes 163Bookmarks 60
TACIXAT@TACIXAT
@NigelHiggs7 TU Delft has some research on this, training with domain randomization on the drone configs allows it to transition to a real drone pretty well https://arxiv.org/pdf/2504.21586
that's what I'm working on now for AI Grand Prix, first my sim to their sim, then to real for the finals
7hViews 1.6KLikes 30Bookmarks 16
TACIXAT@TACIXAT
@mSanterre I wrote it from scratch, JAX for drone kinematics and Raylib Python bindings for visualization
7hViews 1.3KLikes 46Bookmarks 10
kache@yacineMTB
just FYI
7hViews 2.5KLikes 19Bookmarks 2
kache@yacineMTB
That also pass sim2real
7hViews 2.7KLikes 19Bookmarks 1
Nigel Higgs@NigelHiggs7
@TACIXAT That’s sick although, how well does this generalize to a real drone situation?
7hViews 1.8KLikes 2
Henry Paradiz@0xtechnexus
@TACIXAT yea but would you strap a 3090 to a drone and blow it up
9hViews 627Likes 1
max@mSanterre
@TACIXAT Damn son that's cracked
7hViews 387Likes 10
Nicholas Perry@RegolithHunter
@TACIXAT And you can also make custom modules that can give you significant speed increases. I hit 9100x compute speedup just by prompting an optimized module.
Creativity can often beat intelligence.
7hViews 797Likes 1Bookmarks 1
max@mSanterre
@TACIXAT Which sim is this?
7hViews 1.4KLikes 3
Ashish kumar Singh@ashishkmr472
@TACIXAT Looks great, what algorithm did you use? I built something similar on my 4080 too, pure fpv mode with just stick controls - https://github.com/AshishKumar4/FlyDreamer
4hViews 179Likes 4
Rev. Poppy Haze 𓅋@poppy_haze
@TACIXAT the important bit is to get it to run on a ruggedized Intel Atom, and CF card that costs the taxpayer $200 each
6hViews 126Likes 4
Muh Tweeter@muh_tweeter
@NigelHiggs7 @TACIXAT this is also a pretty interesting paper for positional controllers https://arxiv.org/html/2509.11481v1 I believe it's even upstreamed into px4 now if you build it correctly
6hViews 19Likes 1Bookmarks 1
name@exigenced
@TACIXAT Their compute constraint is coming from inside the house
7hViews 438Likes 5
Azide@evilmoderazide
@TACIXAT i hate how "useful experimentation" now usually just means llms
7hViews 477Likes 4
N@Mirim08166421
@TACIXAT Do you think this has any advantage over algorithmically maneuvering the drone?
5hViews 306
JS@spoozpriest
@TACIXAT also like the gymnasium framework is 100% right there and it accomplishes insane shit off a 10 year old mac cpu.
7hViews 544Likes 3
CommonKnowledge@KnowCommon
@0xtechnexus @TACIXAT You train on a 3090 and quantize the model so that it can run on a smartphone
7hViews 32Likes 1
Fuggy@CEOofFuggy
@yacineMTB I could do this without any AI
6hViews 28Likes 1
GoldenGreen@GoldenGreen_
@TACIXAT I also like how people talk exclusively through terms that have been effectively obliterated by marketing firms (compute) instead of just referring to flops - which compute used to mean, but now effectively doesn't, because it isn't a commodity in this context
6hViews 777Likes 2
Load more posts