From 4e84671a15bb0f8ca70b5a0b5906a41f8aec0fb5 Mon Sep 17 00:00:00 2001
From: Taylor Howell <taylorhowell@google.com>
Date: Fri, 5 Dec 2025 02:07:07 -0800
Subject: [PATCH] MuJoCo Warp documentation: Determinism.

PiperOrigin-RevId: 840630917
Change-Id: Ie5f3ac34cee8812f20d6ae4fa6004385c5b048e0
---
 doc/mjwarp/index.rst | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/doc/mjwarp/index.rst b/doc/mjwarp/index.rst
index 6ef6ec01..11a45c92 100644
--- a/doc/mjwarp/index.rst
+++ b/doc/mjwarp/index.rst
@@ -418,6 +418,15 @@ for details and
 `mjlab distributed training <https://github.com/mujocolab/mjlab/tree/main/docs/api/distributed_training.md>`__ for a
 reinforcement learning example.
 
+**Is MJWarp on GPU deterministic?**
+
+No. There may be ordering or *small* numerical differences between results computed by different executions of the same
+code. This is characteristic of non-deterministic atomic operations on GPU. Set device to CPU with
+``wp.set_device("cpu")`` for deterministic results.
+
+Developments for deterministic results on GPU are tracked in this
+`GitHub issue <https://github.com/google-deepmind/mujoco_warp/issues/562>`__.
+
 **How are orientations represented?**
 
 Orientations are represented as unit quaternions and follow :ref:`MuJoCo's conventions<siLayout>`: