Nicolas Brusco

400 posts

Nicolas Brusco banner
Nicolas Brusco

Nicolas Brusco

@Brusco_RF

Software sucks! Write hardware. #FPGA

Scottsdale, AZ Katılım Nisan 2022
548 Takip Edilen252 Takipçiler
Nicolas Brusco
Nicolas Brusco@Brusco_RF·
Fair enough. My experience with drone RL is opposite, SAC produced good results while PPO did some weird stuff (it liked to tumble… not sure why) Definitely a steeper learning curve for SAC though. Lot more places to shoot yourself in the face, and there’s a million papers on ways to tweak it that you can get lost in for weeks
English
1
0
0
114
Joseph Suarez 🐡
Joseph Suarez 🐡@jsuarez·
@Brusco_RF @natolambert I tried it, was terrible. From my work last summer, I don't think that algo has anything going for it at all (read the bootstrap article) and it's mostly that robotics people have just spent forever tuning it on their problems
English
1
0
4
313
Nathan Lambert
Nathan Lambert@natolambert·
I was not ready for this PPO vs GRPO debate. Here we go again. The truth is just that policy gradient good.
English
18
14
593
51.2K
Nicolas Brusco
Nicolas Brusco@Brusco_RF·
@jsuarez @natolambert You ever mess around with SAC? I know puffers whole thing is massive SPS which removes SAC’s main benefit (sample efficiency) but I’m still curious your thoughts. I default to SAC for robotics control problems
English
1
0
2
333
Joseph Suarez 🐡
Joseph Suarez 🐡@jsuarez·
@natolambert It's a stupid debate. GRPO is a bad algorithm, used because LLM people can't train value functions. Try GRPO in non-llm RL and it won't be close
English
7
0
81
5.4K
Modal Motors
Modal Motors@ModalMotors·
@Brusco_RF @DetroitTechGuru @Rich_Pluta It is a new architecture. Transverse flux. It is superior to radial flux if you can get it to work. We maximize the magnetic flux. Radial flux loses about 40-55% of the magnetic flux. We don’t lose any.
English
1
0
2
11
Rich Pluta
Rich Pluta@Rich_Pluta·
Just ran into Modal Motors. Non-rare earth electric motors for drones and robotics. Similar costs to Chinese motors, all U.S.-sourced. Incredible.
Rich Pluta tweet media
English
81
215
3.5K
303.4K
Nicolas Brusco
Nicolas Brusco@Brusco_RF·
@odayHQ Where’d u hear about the leaderboard? I hadn’t heard that
English
1
0
1
35
Oday
Oday@odayHQ·
@Brusco_RF Great work. I did 15s with neural but had the condition of the camera facing forward for gate detection later. There will be a leaderboard for VQ2 so that's exciting. Expect to see you at the top.
English
1
0
2
20
Nicolas Brusco
Nicolas Brusco@Brusco_RF·
When I first flew this course manually I thought 15s would be a good goal for the RL. It just broke 5s! #AIGrandPrix #Anduril
English
6
2
17
3.5K
Mickabout
Mickabout@DetroitTechGuru·
@Brusco_RF @Rich_Pluta @ModalMotors Data sheet is coming. To summarize here we are typically 20-30% lower mass at the same torque and with much higher peak efficiency and more importantly Average Efficiency over the full load/speed range. This translates into longer flight time/runtime for any DC powered product
English
1
0
2
44
Nicolas Brusco
Nicolas Brusco@Brusco_RF·
@wavefnx I do a similar thing with a tmux tab for codex and claude. Except I make them communicate via .md files. How'd you make them talk directly?
English
0
0
1
42
wavefnx
wavefnx@wavefnx·
I've created the first multi-model cross-provider review system, and you're blackpilling? this is now a 200 quintillion param model Jokes aside, it's been using using them co-op and works natively with each providers CLI not only w my TUI They are stricter with each others code
wavefnx tweet media
wavefnx@wavefnx

Claude Opus 4.7 and Codex 5.5 work well with each other. I built them a system a while ago so they do their own thing and if the system is built in a good way, it can have very good results When done, they play a little sound That's what they were missing

English
3
0
15
1.1K
Nicolas Brusco
Nicolas Brusco@Brusco_RF·
@KinvertOG Incredibly cool man. I wonder what the sim2real gap is like for applications like this
English
1
0
1
55
Kinvert
Kinvert@KinvertOG·
Sound on. Bat finding the bug using only the info of its last chirp, and the FFT of the reflections. One cool thing about PufferLib, you can train deterministically. So I could train again, get the exact same result, and draw what the bat sees. x.com/KinvertOG/stat…
Kinvert@KinvertOG

PufferLib env I made, a bat chasing a bug. Sound on. It sends out chirps and it's only observations are FFT for each ear catching reflections, the last chirp it made, and a timer. it senses the doppler shift in the bug on its own. PufferLib one shot this.

English
4
5
27
2.9K
Nicolas Brusco
Nicolas Brusco@Brusco_RF·
@CalebisGross Interesting. Seems like a poorly designed filter to me. Anyway, points moot now
English
1
0
2
17
Caleb
Caleb@CalebisGross·
@Brusco_RF That's odd. Been doing the same work and never had an issue. Definitely wasn't using 'safe' language either. Mentioned targeting systems, countermeasures, stealth, etc.. was actually surprised I didn't get dinged considering how many posts I've seen about it.
English
1
0
1
47
Nicolas Brusco
Nicolas Brusco@Brusco_RF·
I really really wanted to like Fable 5 but it’s extremely frustrating that it lobotomizes itself to Opus every other prompt. Maybe that’s just the reality of working on RL for drones.
English
3
0
2
184
Nicolas Brusco
Nicolas Brusco@Brusco_RF·
I’ll be ten prompts deep when I notice it’s doing something stupid. > look inside > it’s Opus
English
1
0
1
117
Nicolas Brusco
Nicolas Brusco@Brusco_RF·
@BowTiedNiners They say if you’re good enough at your job, you get to do other people jobs too! I’ve run into this issue many times. It sucks.
English
1
0
1
85
AeroNiners 🇺🇸
AeroNiners 🇺🇸@BowTiedNiners·
They saw I’m competent so they are pulling me to support other teams Guess I’m a flight test engineer now
English
7
2
73
2.9K
Nicolas Brusco
Nicolas Brusco@Brusco_RF·
@yacineMTB @internetvin I’m curious, what control rates do you generally use? And what physics tick? I’ve found that training is highly sensitive to both.
English
0
0
1
146
kache
kache@yacineMTB·
today; the goal is to reduce the time of training for this by 10x. This took 27.8 minutes to train at 200m steps 120k steps per second. reproduction from mujoco playground using mjwarp
English
29
12
480
27.3K
Nicolas Brusco
Nicolas Brusco@Brusco_RF·
VQ1 gives full perfect world information. Perfect location, velocity, orientation, gate positions, etc. That makes vision unnecessary, which was a bummer, because I spent a long time on my vision models and EKF. VQ2 should noise the data and remove some fields entirely, so more fun And yeah I also thought the 20 degree camera was weird. I actually emailed them about it. I used to race with a 45 degree camera, most racers go even steeper.
English
0
0
2
46
TACIXAT
TACIXAT@TACIXAT·
@Brusco_RF hell yea, is this flying off vision? I don't think the 20 degree camera angle will support much faster than 12s with keeping gates in view, haven't started training yet though
English
1
0
1
62
Nicolas Brusco
Nicolas Brusco@Brusco_RF·
@RFreshhead @BishopAerospace Andruil fixed the camera angle for this competition at 20 degrees, which I thought was silly. When I used to race drones my angle was 45, most people go even steeper
English
0
0
2
42
Nicolas Brusco
Nicolas Brusco@Brusco_RF·
Couple caveats worth mentioning. The VQ1 competition gives perfect world information. The controller always knows the drone’s exact location, orientation, velocity, and where the gates are. That’s what makes what you’re seeing possible. VQ2 will (likely) give noisy raw sensor data. It will need to be supplemented by vision. That means the drone will need to point the camera towards the gates and it will be slower and look less impressive. It’ll be more fun though! The real world is messy, our algorithms need to clean it up. That’s the real competition
English
1
0
1
43
Nicolas Brusco
Nicolas Brusco@Brusco_RF·
@BishopAerospace This is RL. It’s way too fast to fly manually, the camera isn’t even pointed at the gates anymore
English
1
0
2
104

Keşfet

@natolambert @jsuarez @ModalMotors @DetroitTechGuru @Rich_Pluta @odayHQ @Mark_o1t @wavefnx