Not sure why this is hitting the home page right now but people may also be interested in Mujoco Playground [1] which is the latest RL environment wrapper of mujoco, implementing both classic deepmind-control benchmarks, and some very new interesting ones!
The physics engine I'm using is called MuJoCo. And if you're wondering why I didn't write my own physics engine, it's basically because I don't have 20 years.
It's what put MuJoCo on my radar recently! But I was surprised to not see him do any kind of gradient descent to optimize his hyperparameters. MuJoCo has a JAX backend so it should be fairly straightforward.
I'm pretty sure he has used gradient descent in previous videos to optimize systems, maybe this time it was just easier to hand tune rather than set up an optimization feedback harness around MuJoCo.
Though he had to resort to manual calibration. I always find he has interesting problems for domain experts and would like to see him team up with one. Also with programmers for faster programs than self taught python.
In the video in question, he doesn't seem able to choose a good scoring function for the stochastic solver (even over multiple weeks), seemingly choosing a linear sum of distances (see 8:50) between simulation and reality. That's a mistake that not even an undergraduate should make. He needs some domain experts.
While we have people's attention here, we're also working on an official browser-based interactive viewer that allows people to specify a URL to their own model and share a link.
We are using MuJoCo to train a G1 humanoid robot right now. The best thing is that we do not need to fight with NVIDIA software and that it runs on macOS.
This makes me so happy and excited! Often my mind wanders into the unknown, imagining what would happy to X if it did this? Would it have friction, etc?
I am looking forward to a way I can easily describe a scenario and have an LLM build a legitimate simulation for it. No more hypothetical talk! Next best thing to actual experimentation (can be a useful tool in convincing others to join you/support you in said real experiment - “see? I tested it in a simulation and it behaves exactly that way! We need to try this..”).
One thing to be mindful of is that you can get a simulation to behave in (almost) any way you want if you set the parameters right, so you should take care to understand the assumptions that you're baking into your sim before taking its results as gospel.
Build things yourself. Using LLMs doesnt help you understand anything, they will just give you an annoying case of dunning kruger. Using them will only make you retart-d
Not sure why this is hitting the home page right now but people may also be interested in Mujoco Playground [1] which is the latest RL environment wrapper of mujoco, implementing both classic deepmind-control benchmarks, and some very new interesting ones!
[1] https://playground.mujoco.org/
Probably because of the StuffMadeHere video
This is what StuffMadeHere used in his latest video to simulate a mini-golf course! https://www.youtube.com/watch?v=2OfjZ3ORJfc&t=368s
The physics engine I'm using is called MuJoCo. And if you're wondering why I didn't write my own physics engine, it's basically because I don't have 20 years.
It's what put MuJoCo on my radar recently! But I was surprised to not see him do any kind of gradient descent to optimize his hyperparameters. MuJoCo has a JAX backend so it should be fairly straightforward.
He is much better at building hardware than he is writing software.
He seems pretty damn good at both.
I'm pretty sure he has used gradient descent in previous videos to optimize systems, maybe this time it was just easier to hand tune rather than set up an optimization feedback harness around MuJoCo.
That the calibration got the simulation so close to reality was quite impressive.
Though he had to resort to manual calibration. I always find he has interesting problems for domain experts and would like to see him team up with one. Also with programmers for faster programs than self taught python.
The guy has a masters in compsci and mechanical engineering, he has done both python and c++ afaik for his projects
In the video in question, he doesn't seem able to choose a good scoring function for the stochastic solver (even over multiple weeks), seemingly choosing a linear sum of distances (see 8:50) between simulation and reality. That's a mistake that not even an undergraduate should make. He needs some domain experts.
I did comp sci, I sucked at this stuff, and I have a masters.
Low level stuff? Os? Distributed systems? Multithreaded code? Now that’s more on my alley
I think the parent comment is specifically addressing why the black box (or stochastic?) optimizer he used not working.
While we have people's attention here, we're also working on an official browser-based interactive viewer that allows people to specify a URL to their own model and share a link.
Very early experimental prototype at the moment, doesn't work super well on phones yet, but you can try it at https://saran-t.github.io/mujoco/?model=https://raw.githubus...
It'll be relocated to live.mujoco.org over the next few days.
We are using MuJoCo to train a G1 humanoid robot right now. The best thing is that we do not need to fight with NVIDIA software and that it runs on macOS.
PS: I just finished a first draft for agentic skills around working with MuJoCo in Python. Feel free to check them out here: https://github.com/prathje/agentic_mujoco_skills
My two cents, I actually worked with getting MuJoCo's physics gradients to train quadruped locomotion policies. Really interesting field.
https://github.com/saucesaft/differential_policies/
Mujoco is also key part of nvidias Newton physics system
https://github.com/newton-physics/newton
MuJoCo is great. I have it running in the browser for robotics simulation. See for example https://visibot.com/sheet/examples/humanoid_walking.v
People have made cool racing education simulators with this too: https://github.com/FT-Autonomous/ft_grandprix.
This makes me so happy and excited! Often my mind wanders into the unknown, imagining what would happy to X if it did this? Would it have friction, etc?
I am looking forward to a way I can easily describe a scenario and have an LLM build a legitimate simulation for it. No more hypothetical talk! Next best thing to actual experimentation (can be a useful tool in convincing others to join you/support you in said real experiment - “see? I tested it in a simulation and it behaves exactly that way! We need to try this..”).
One thing to be mindful of is that you can get a simulation to behave in (almost) any way you want if you set the parameters right, so you should take care to understand the assumptions that you're baking into your sim before taking its results as gospel.
> what would happy to X if it did this? Would it have friction, etc?
MuJoCo does not simulate at a scale where you can discover whether something has friction or not. Friction is a parameter to MuJoCo's simulations.
Build things yourself. Using LLMs doesnt help you understand anything, they will just give you an annoying case of dunning kruger. Using them will only make you retart-d
Wait what's new ? This is an established library
[dead]
[dead]