MJX H1 --> Menagerie?

[kevinzakka]

Hi!

Just wanted to say congrats on the amazing paper (real world results are always so exciting to see!) and the beautiful visualizations! I wanted to check in with y'all to see if you would you be willing to submit a PR to Menagerie with the MJX variant of the H1? Would be incredibly helpful to the community.

[kevinzakka]

Similarly, if you have a better MJX-variant of the Go2, I'd love for you to contribute it as well!

[jc-bao]

Hi @kevinzakka, thank you so much for the kind words! We're thrilled that you liked the paper and visualizations. We'd be happy to contribute to Menagerie. Regarding Go2, we’ve identified a significant sim-to-real gap in torque control, largely due to incorrect damping settings in the original Menagerie repo. Additionally, we’ve greatly simplified the contact model by keeping only the four feet in contact during real-world experiments to accelerate simulation. This raises some concerns about updating the Go2 MJX model, but we’re certainly open to contributing a version with the corrected damping model.

As for H1, we’ve ported it from Menagerie and are currently conducting further sim-to-real validation. Once the verification is complete, we’ll contribute a PR for the MJX variant.

We’re looking forward to contributing and supporting the community!

[HaoruXue]

Hi @kevinzakka thanks for liking our work!

A quick question on the H1 MJX model - since all the contact geometries used by the H1 model in Menagerie is already compatible with the latest MJX, I don't see why not just use the H1 already in there for MJX?

In other words, what would you guys be looking for in terms of a "MJX variant of the H1"? Like just more simple contacts?

[kevinzakka]

@HaoruXue I didn't inspect the model in depth but assumed you had done optimizations (e.g. solver parameters) specific to MJX!

[HaoruXue]

@HaoruXue I didn't inspect the model in depth but assumed you had done optimizations (e.g. solver parameters) specific to MJX!

I see, and yes in deed! Then let us test internally a little more as @jc-bao mentioned above, and then we can be ready to contribute back!