Robotics builder Yacine compiles MuJoCo into CUDA graphs using NVIDIA Warp, prompting debate over whether standard C is fast enough
Launching CUDA graphs from C code enables highly accelerated simulations
Users are excited about MuJoCo Warp compiling simulations as CUDA Graphs because it enables much faster robotics training and breakthroughs like acoustic physical simulations.
Most Activity
@yacineMTB
@yacineMTB you don't need cuda for any of this. It will be plenty fast in C
You can use mujoco warp to compile your simulation as a cuda graph, and then launch that cuda graph from C to have absurdly fast robotics simulations. Nvidia warp also open sources a bunch of really great renderers, fluid sim stuff
@yacineMTB i fully paralleled the acoustics/audio physical simulation using warp and trained a lot of locomotion/navigation policies conditioned on sound. all the policies can be trained in 5 minutes using 3090. i think i can generate 100 icra papers now.
@york1to lets do sim2real. do you think we could build a little ultra sound harness and use that ? its ray tracing and bouncing around, i think you could easily do this sim2real
@york1to if you know what frequency you're sending and receiving; and build a DIY phase microphone, i'm sure you could make it work with just simulating it geometrically
@yacineMTB i mean really, for real, a complete breakthough in acoustic robotics
@york1to this is more possible than you think, if you're joking. it's so possible that i believe you
@york1to @yacineMTB yacine & co making a feathered echo locating predatory bird with sim2real RL and we are only halfway through 2026
@york1to post video please
@yacineMTB there are two ways of simulating, geomtrically is ray tracing+diffractions. when it comes to low frequency bands, we need to solve some kind of wave based equations
@york1to it's just ray tracing
@york1to like the cheap ultrasound stuff you can get from aliexpress
@york1to holy shit
@jsuarez some of the stuff that i want to does need parallel, like fluid sim etc
@york1to @yacineMTB So cool
@york1to @yacineMTB omg
@york1to really the way i would do this is build a harness in real life; collect data, then work on simulator using warp for parallel. btw what training loop are you using for this?
@yacineMTB Damn it, stop mentioning cool stuff. I need to finish the other things Im working on
@yacineMTB Why dibt games use that fluid sim ?
@yacineMTB one of my labmates is doing that exact research so i'm not doing that lol mic arrays and mics on ee acting like tactile sensors is what i am focusing on