Yacine claims the same policy trains from pixels in two minutes
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)#440
idk you can train drones to fly on a 3090 in a couple of hours, this is a 3x64 nn
It is kinda wild that one essentially needs ~11k of compute a month (an 8xH100 box) to perform any form of useful experimentation, and that's the lower bar.
Never in my life have I felt similarly compute-constrained.
Positive users praise small neural nets training drone policies from pixels on a single RTX 3090 as a refreshing alternative to large models, while negative users dismiss the approach as unnecessary or inferior to traditional programming.
I can train the same policy in 2 minutes, policies that are trained off of pixels
idk you can train drones to fly on a 3090 in a couple of hours, this is a 3x64 nn
@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
@mSanterre I wrote it from scratch, JAX for drone kinematics and Raylib Python bindings for visualization
just FYI
That also pass sim2real
@TACIXAT That’s sick although, how well does this generalize to a real drone situation?
@TACIXAT yea but would you strap a 3090 to a drone and blow it up
@TACIXAT Damn son that's cracked
@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.
@TACIXAT Which sim is this?
@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
@TACIXAT the important bit is to get it to run on a ruggedized Intel Atom, and CF card that costs the taxpayer $200 each
@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
@TACIXAT Their compute constraint is coming from inside the house
@TACIXAT i hate how "useful experimentation" now usually just means llms
@TACIXAT Do you think this has any advantage over algorithmically maneuvering the drone?
@TACIXAT also like the gymnasium framework is 100% right there and it accomplishes insane shit off a 10 year old mac cpu.
@0xtechnexus @TACIXAT You train on a 3090 and quantize the model so that it can run on a smartphone
@yacineMTB I could do this without any AI
@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
Yacine claims the same policy trains from pixels in two minutes
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)#440idk you can train drones to fly on a 3090 in a couple of hours, this is a 3x64 nn
It is kinda wild that one essentially needs ~11k of compute a month (an 8xH100 box) to perform any form of useful experimentation, and that's the lower bar.
Never in my life have I felt similarly compute-constrained.