Albert Li

Any chance this could get merged soon? I'd also be interested in a backport at least through Humble.

+1 on wanting this feature!

I'm not sure - the bug I'm seeing here is more like an "off by 1" situation where the first (inferred) dummy joint that fixes the base of the robot...

> So I think it's quite clear. If we want to model fixed stuff, like Franka panda, all the geoms from link0 or aka. base, , should not be wrapped...

@baixianger See below. All the assets for this MJCF can be found [here](https://github.com/deepmind/mujoco_menagerie/tree/main/franka_emika_panda). ``` ```

Hi @gomezzz, I haven't checked whether it works for earlier versions. [This commit](https://github.com/google/jax/commit/c90f1f0c960da87716d5c49e02efc4571e8a0781) from 6 months ago seems relevant if you want to control versioning, though.

Hi @gomezzz, my bandwidth is very limited in the next couple of weeks, so I probably won't be able to write this PR. I think separating the testing issues into...

@blurgyy Thanks for the edits/instructions - I finally got back around to looking at this and successfully built `jax-tcnn`. I have a follow-up question: my goal is to train a...

> Can you add the additional cost term One wrinkle here is that `total_return` generally includes penalties against actuation as well. In path integral control theory, the cost `q` should...

> Let's add the cost term from the algorithm and default γ to zero. But then when `γ` is nonzero and we have actuation costs in `total_return` the result still...