Memory leak in mjs_delete when detaching a body

[chengruiz]

Intro

Hi!

I am a PhD student at Zhejiang University. I use MuJoCo for my research on legged locomotion.

I use massively parallel environments for reinforcement learning with procedural terrain generation. After upgrading MuJoCo from 3.3.3 to 3.3.4 (which version replaces mjs_detachBody with mjs_delete), I found that a continuous memory leak occurs that had never happened before. I checked the source of MuJoCo and found that in version 3.3.3, mjs_detachBody calls model->Detach(body);, where subtree->Release(); is invoked.

https://github.com/google-deepmind/mujoco/blob/f978b4eea368450267dd5e8825347653bec7355e/src/user/user_model.cc#L844C1-L850C1

However, after version 3.3.4, the implementation is moved to void mjCModel::operator-=(mjsElement* el), where the subtree->Release(); is removed. If I manually add subtree->Release();, the memory leak is fixed.

https://github.com/google-deepmind/mujoco/blob/8de86c9944e17efd04a3972a630a79a7da30736e/src/user/user_model.cc#L750C1-L757C6

Do I use the api in the wrong way? Or is this indeed a bug?

My setup

MuJoCo 3.3.3 - 3.3.6 C version

What's happening? What did you expect?

Memory leak.

Steps for reproduction

Simple code for reproduction may be hard to provide.

Minimal model for reproduction

No response

Code required for reproduction

No response

Confirmations

• I searched the latest documentation thoroughly before posting.
• I searched previous Issues and Discussions, I am certain this has not been raised before.

[chengruiz]

I rechecked the source code and found that the detached elements are moved to detached_ and deleted when the mjCModel is deconstructed. Although it might be reasonable, for my application where bodies are frequently attached and detached, MuJoCo does not provide an interface to clean the detached elements.