I unfortunately can't use any anthropic models for this, because there is no way for me to know whether I'm silently getting nerfed & sandbagged. Not allowed to make the world a better place

Not enough that they're closed, open source devs need to be slowed down too

My suspicion is that the contact dynamics, code, is not at all optimized for ultra fast RL. The only reason it is as slow as it is is because no one really tried, and I'm going to try

This is all going to be open source. It's a project. My goal is to accelerate robotics

The problem with the "frontier" distinction is that it is laughably easy to be frontier - there are so many green fields not yet explored. It's like saying you're at the frontier in the middle of kansas. There's fertile ground everywhere.

@ChaseBrowe32432 silently giving people horse shit information is poisoning the well for everyone. why don't they make it clearly refuse?

Particularly, reinforcement learning for robotics, that anyone can do with a single GPU at home. It's going to be CUDA only (thanks jensen..). I am using 3x 4090s, but each training baseline will run on a single 4090.

0 python will be used

I'm going to base my work on top of pufferlib and follow along their development. They are scarily good at making environments for all kinds of tasks - I've done remarkable things with their RL loop. They're taking contracts if you think you have a problem they can help with

@yacineMTB this guy walkin dumb as hayl

@yacineMTB @ChaseBrowe32432 active deception is the key issue here that people keep missing, and i have to wonder if they're missing the point on purpose

@yacineMTB bro lets move the families to shenzhen for a few years

@LouisLeLay4 Yes! I'm going to follow in homemadegarbage's foot steps and make home made garbage move. The goal is to give people a reproducible RL loop they can use

@yacineMTB Will you deploy on real robot afterwards?

@yacineMTB Wouldn't a better sim randomize some elements of the physics parameters so that your model learns to adapt to differences in physics so when it gets to the real world it just works?

@yacineMTB Convex decomposition sucks. But probably not going to be an issue for basic humanoid stuff. Maybe could approximate some collisions with spheres and make it go faster than convex shapes would.

Race one vs a “ghost time”.

“ghost” = moving average of the average time per section of the race track. Reward is distance ahead of ghost applied for the time it sustains lead. Negative if behind.

Robustify it… Add uneven surfaces and barriers or track and penalize for how much it strays from center.

@yacineMTB you go to a restaurant order a ribeye, but chef swaps it for the cheapest cut because it's good for you, but still pay for a ribeye, and the kitchen has tinted windows

@yihyunCS @yacineMTB really? they could have easily done this without notifying anyone, and no would would have ever known. they mentioned that they're doing this in the spirit of honesty

@yacineMTB You're cooking

@yacineMTB I love how when he falls only the feet have collision bounds.

@yacineMTB If you only simulate the foot/ground contact, the sparsity structure of the constraint hessian, i.e JtDJ, is the same as the sparsity structure of the JSIM. As far as I'm aware none of the official mujoco implementations exploit this fact.

@yacineMTB I can already train locomotion very very fast 200K sps is fast but it changes nothong. If you could 10x the speed, maybe thats relevant. But even then, this isnt where help is needed.

Where we need help is in high contact physics envs.