From ab939683bc2c8570eacbe1d53d03a24eb6a3eb6a Mon Sep 17 00:00:00 2001
From: Taylor Howell <taylorhowell@google.com>
Date: Thu, 8 Jan 2026 05:33:25 -0800
Subject: [PATCH] MuJoCo Warp documentation: Why are numerical results from
 MJWarp and MuJoCo different?.

PiperOrigin-RevId: 853696967
Change-Id: I0cf8539e341d2aa1cfd07ac75de18128ce9beb3b
---
 doc/mjwarp/index.rst | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/doc/mjwarp/index.rst b/doc/mjwarp/index.rst
index e2e3a608..c95f79e2 100644
--- a/doc/mjwarp/index.rst
+++ b/doc/mjwarp/index.rst
@@ -498,6 +498,14 @@ size ``Data.nworld``. Some :class:`mjw.Model <mujoco_warp.Model>` fields having
 Additionally, certain fields including ``Model.qM``, ``Data.efc.J``, and ``Data.efc.D`` are padded to enable fast
 loading on GPU.
 
+**Why are numerical results from MJWarp and MuJoCo different?**
+
+MJWarp utilizes `float <https://nvidia.github.io/warp/modules/functions.html#warp.float32>`__s in contrast to MuJoCo's
+default double representation for :ref:`mjtNum`. Solver settings, including iterations, collision detection, and small
+friction values may be sensitive to differences in floating point representation.
+
+If you encounter unexpected results, including NaNs, please open a GitHub issue.
+
 Compilation
 -----------