I've gotten a mujoco sim RL training loop for a unitree robot at 200k SPS. I'm looking into the physics for friction, contact dynamics. My goal: can I reproduce & beat the mujoco playground RL baselines
This is running in my web browser with raylib. Its the baseline
Positive users praise the impressive 200k SPS speed and realistic simulation details in the browser MuJoCo RL training for the Unitree robot, while negative users criticize AI companies for limiting paid product capabilities.
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
I've gotten a mujoco sim RL training loop for a unitree robot at 200k SPS. I'm looking into the physics for friction, contact dynamics. My goal: can I reproduce & beat the mujoco playground RL baselines
This is running in my web browser with raylib. Its the baseline
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
I've gotten a mujoco sim RL training loop for a unitree robot at 200k SPS. I'm looking into the physics for friction, contact dynamics. My goal: can I reproduce & beat the mujoco playground RL baselines
This is running in my web browser with raylib. Its the baseline
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.
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
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 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
@yacineMTB You're cooking
@yacineMTB I love how when he falls only the feet have collision bounds.
@yacineMTB 200k SPS is impressive already
@yacineMTB Man, all they're trying to do is add shareholder value.
@ChaseBrowe32432 @yacineMTB Ah yes so we should just accept that they SELL a product you PAY for that intentionally sabotages you without you knowing in some cases? Keep in mind Claude is also heavily targeted at enterprises. It’s not a free private use product where this might be more defensible.
@yacineMTB You're pretty good at analogies. I like this one a lot.
@lyc_aon @yacineMTB This part is a bit too on the nose 😂
How it feels to try to do some perfectly fine future shit and the model says no
I started working on this a while back. but since time is money, I was broke, if this can be solved, it will absolutely scale all your robotic research: 🎁
CoVeRT: Constraint Valid Environment Resampling and Tracking for synthetic environment generation
The system is basically a validity layer for generated simulation states.
Start with one working task scene. It extracts the contact graph, support graph, reachability bounds, joint limits, object affordances, task constraints, collision margins, closest point distances, and penetration depth. Then it generates new scenes by changing poses, clutter, friction, mass, geometry, dynamics, or layout, but only accepts states that still obey the world rules and the task.
So one clean setup can become thousands of usable states: curricula, edge cases, near failures, labels, validity scores, and rejection reports for the broken ones.
This applies beyond robotics. Any 2D or 3D simulation where an AI learns from generated worlds.
Plug that into your harness and have it do some research.
@yacineMTB If they don't nerf stuff their govt contracts will be at risk. The Trump admin is no joke
@yacineMTB Me: scrapping excel files with Claude 🤡
@yacineMTB What are you using the run the simulations? Do you have a custom framework you are working from or an off the shelf tool for it?
@yacineMTB And thats just beginning. As they grow and control more expect even more of this crazed attitude.
Thats why need more support for open models. Need some actual foundation for training and improving open weight models.
@yacineMTB i look forward to seeing this. i am bored and need a challenge
I've gotten a mujoco sim RL training loop for a unitree robot at 200k SPS. I'm looking into the physics for friction, contact dynamics. My goal: can I reproduce & beat the mujoco playground RL baselines
This is running in my web browser with raylib. Its the baseline
