Nicolas Brusco

Software sucks! Write hardware. #FPGA

[Nicolas Brusco @Brusco_RL]

@TaoCTakahashi Double check all your axies, and plot the observation inputs into your network. That should make the bugs stand out. At least that’s where mine were hiding

[Nicolas Brusco @Brusco_RL]

I'm now thinking the VQ2 floor might be around 14s. Initially I thought 12s. But it has improved so little in days of training, any improvement in speed just runs up against degraded localization. This may be more difficult than I believed.

[Nicolas Brusco @Brusco_RL]

@HenrySpeiser @anduriltech There are very sample efficient methods out there. 30-120 SPS is plenty if you've got a few days to train

[Nicolas Brusco @Brusco_RL]

LIKE A BAT OUT OF HELL!
@Anduriltech
#TheAIGrandPrix

[Nicolas Brusco @Brusco_RL]

@spiffyml Whats your vision stack? I used YOLOv8

[Nicolas Brusco @Brusco_RL]

@spiffyml @anduriltech Ground truth position is not provided in VQ2, so you do need a localization stack

[Nicolas Brusco @Brusco_RL]

@famgod2022 @anduriltech It is trained for this track on this sim only. For the physical qualifier I will train a general model, already working on it

[Nicolas Brusco @Brusco_RL]

@10_X_eng @anduriltech Badass!

[Nicolas Brusco @Brusco_RL]

@kubasienki @anduriltech I presume so, but it's all learned behavior so no way to know for sure.
20 degrees still irks me. Real racing drones use like 45 - 70 degrees. It could have been so much faster

[Nicolas Brusco @Brusco_RL]

@BishopAerospace @anduriltech Thank you!
I'll be happy to explain exactly how I did it.... Right after I get the money!

[Nicolas Brusco @Brusco_RL]

@odaybrahem I never had much luck with sim2sim

[Nicolas Brusco @Brusco_RL]

Since Andruil has't released the leaderboard yet, has anyone else completed VQ2? I'd like to compare notes

[Nicolas Brusco @Brusco_RL]

@kubasienki End of this month. But theres a chance they will extend it. I don't think as many people have completed it as they hoped

[Nicolas Brusco @Brusco_RL]

My UTD is dynamic, the sim runs in a separate process from the training. Separate machine, actually. They both run as fast as possible, 6 is just the balance they strike. I have a hard cap at 10, SAC gets weird after that.

And my plan for the physical qualifier is.... no plan. I've been fully focused on getting through the virtuals first. Guess I should get to work...

[Nicolas Brusco @Brusco_RL]

@BishopAerospace Thanks! I'm on vacation right now, but when I get home I'll have a fresh video of the entire run. Its fast!

[Nicolas Brusco @Brusco_RL]

Training has been hitting a hard wall at the later gates. But I am starting to believe that tonight's the night

[Nicolas Brusco @Brusco_RL]

sim runs at 120sps, I use K=2 for control so training is effectively 60sps. Its not a lot, but for SAC it's plenty. I've always been an SAC guy, never liked PPO. Never tried model-based like Dreamer.
Now that I've worked out all the bugs, training takes ~6 hours from scratch to course completion on a single RTX3090. I run UTD ~6

[Nicolas Brusco @Brusco_RL]

@kubasienki Oh. I don't do that at all. i train directly in Andruils sim

[Nicolas Brusco @Brusco_RL]

@kubasienki What do you mean by system ID?

[Nicolas Brusco @Brusco_RL]

Moving to rate control means building your own rate controller. I recommend cloning betaflight and telling your agent "implement a rate controller identical to betaflight in this sim". That will avoid a lot of pain. Note that the D term will probably sit at 0 in a sim because no control lag. And yaw is a little weird to get right. Fly it by hand to make sure it feels reasonable before you start training. Easy to waste compute on a broken sim, ask me how I know