Saran Tunyasuvunakool

Senior Staff Research Engineer @GoogleDeepMind, Tech Lead for the #MuJoCo Physics Simulator ⚙️ Numericist, Black Hole Physicist, and Mathematician by training.

[Saran Tunyasuvunakool @saran__t]

@7oponaut @_xjdr Please send in a pull request to Menagerie :)

[Saran Tunyasuvunakool @saran__t]

@7oponaut @_xjdr It's unlikely that matmul precision is the issue. Physics simulation can be finicky when it comes to precision requirements, you could try to compile C MuJoCo in single-precision mode and see if the instability is reproducible there.

[Saran Tunyasuvunakool @saran__t]

@yacineMTB @LukeFarritor @luciduous You should be able to implement the PID controller as an actuator plugin: https://t.co/btT07e04NV

[David Blanco-Mulero @DavidBMulero]

Our recommendation? @GoogleDeepMind MuJoCo.
After the update to version 3 the simulation of deformable objects has significantly improved!
It closely matches the groundtruth on both dynamic (left) and quasi-static (right) tasks and allows importing cloth meshes.

[Saran Tunyasuvunakool @saran__t]

@KayraHumanoid Nice :) you probably need to do some work tuning the solver parameters. If you need help feel free to post a question as an Issue on our GitHub repo.

[Saran Tunyasuvunakool @saran__t]

@KayraHumanoid In many robot models you'll want to set joint limits so that parent-child links don't go that near each other anyway, so the filtering becomes a no-op.

[Saran Tunyasuvunakool @saran__t]

@KayraHumanoid You can use connect constraints to model linkages, however the constraints will be "soft" i.e. some violation should be expected like with contacts.

[Saran Tunyasuvunakool @saran__t]

📢 MuJoCo @ CoRL 📢 Following our successful event at ICRA, we are once again hosting a social event in Atlanta on Nov 6th, to coincide with CoRL. Register at https://t.co/PnvFONVhVb (I won't be able to make it though, sadly.)

[Saran Tunyasuvunakool @saran__t]

@BlackHC @j_foerst @FLAIR_Ox @TensorFlow_ That was "in-graph environment" -- move the environment run loop into the TF graph to avoid the GIL. It's still the same environment code at the end of the day.

[Saran Tunyasuvunakool @saran__t]

@LuoYicheng @DanielCHTan97 Thank you for looking into this!

[Saran Tunyasuvunakool @saran__t]

@Stone_Tao @chris_j_paxton @PMOHANSRINIVAS1 I'd presume there's nontrivial overhead coming from pushing data over PCIe vs. keeping everything on device, but the latter doesn't imply that one needs to JIT the simulation and RL together.

[Saran Tunyasuvunakool @saran__t]

@DibblaX @GoogleDeepMind Short answer is yes, but see https://t.co/REsDXbu7Xv for caveats.

[Saran Tunyasuvunakool @saran__t]

@TheSmallQuail @drew_r_hamilton @kevin_zakka Code can be found at https://t.co/KNewLFzoVK

[Saran Tunyasuvunakool @saran__t]

Excited to finally launch this, it's been a pleasure working with @erikfrey and @the_real_btaba to bring GPU and TPU support to MuJoCo.

[Google DeepMind @GoogleDeepMind]

Introducing MuJoCo 3.0: a major new release of our fast, powerful and open source tool for robotics research. 🤖

📈 GPU & TPU acceleration through #JAX
🖼️ Better simulation of more diverse objects - like clothes, screws, gears and donuts
💡 Find out more: https://t.co/KbzxeALTMV

[Saran Tunyasuvunakool @saran__t]

@chris_j_paxton @PMOHANSRINIVAS1 (@erikfrey)

[Saran Tunyasuvunakool @saran__t]

@chris_j_paxton @PMOHANSRINIVAS1 There's a Brax demo that does this, I'm told that it's reasonably performant. https://t.co/dxFbJUrIMF

[Saran Tunyasuvunakool @saran__t]

@chris_j_paxton @PMOHANSRINIVAS1 Nothing stopping you from using jax.dlpack to interop with pytorch, though.

[Saran Tunyasuvunakool @saran__t]

@DavidBMulero @TheSmallQuail We decided to keep both in the release for the time being since they have different trade-offs. (Flex is still much slower, for example)