From 827a989379a1e5672d20b89aaa169d9fd29acb4a Mon Sep 17 00:00:00 2001
From: Antoine RICHARD <antoiner@nvidia.com>
Date: Wed, 30 Sep 2026 15:54:17 +0200
Subject: [PATCH 1/9] Fall back to the env device for RSL-RL runners without
 CUDA

---
 ...ntoiner-fix-rsl-rl-runner-cpu-fallback.rst |  5 ++
 .../isaaclab_rl/isaaclab_rl/rsl_rl/utils.py   | 11 ++-
 source/isaaclab_rl/test/test_rsl_rl_runner.py | 69 +++++++++++++++++++
 3 files changed, 83 insertions(+), 2 deletions(-)
 create mode 100644 source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.rst
 create mode 100644 source/isaaclab_rl/test/test_rsl_rl_runner.py

diff --git a/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.rst b/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.rst
new file mode 100644
index 000000000000..1ed4157a25ea
--- /dev/null
+++ b/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.rst
@@ -0,0 +1,5 @@
+Fixed
+^^^^^
+
+* Fixed RSL-RL training, playback, and export failing on hosts without CUDA, such as macOS, by
+  building the runner on the environment's device when the agent's default ``cuda:0`` is unavailable.
diff --git a/source/isaaclab_rl/isaaclab_rl/rsl_rl/utils.py b/source/isaaclab_rl/isaaclab_rl/rsl_rl/utils.py
index 8505e2b4a7a4..4204f3761ccd 100644
--- a/source/isaaclab_rl/isaaclab_rl/rsl_rl/utils.py
+++ b/source/isaaclab_rl/isaaclab_rl/rsl_rl/utils.py
@@ -12,6 +12,7 @@
 from dataclasses import MISSING
 from typing import TYPE_CHECKING, Any
 
+import torch
 from packaging import version
 from rsl_rl.runners import DistillationRunner, OnPolicyRunner
 
@@ -60,6 +61,9 @@ def create_rsl_rl_runner(
 ) -> OnPolicyRunner | DistillationRunner:
     """Instantiate the RSL-RL runner selected by ``agent_cfg.class_name``.
 
+    The runner uses ``agent_cfg.device``, or the environment's device when that is a CUDA device and
+    CUDA is unavailable (e.g. on macOS), since the agent configuration defaults to ``cuda:0``.
+
     Args:
         env: Wrapped environment the runner trains or evaluates on.
         agent_cfg: Runner configuration.
@@ -68,10 +72,13 @@ def create_rsl_rl_runner(
     Raises:
         ValueError: If the configured runner class is not supported.
     """
+    device = agent_cfg.device
+    if device.startswith("cuda") and not torch.cuda.is_available():
+        device = str(env.device)
     if agent_cfg.class_name == "OnPolicyRunner":
-        return OnPolicyRunner(env, to_dict(agent_cfg), log_dir=log_dir, device=agent_cfg.device)
+        return OnPolicyRunner(env, to_dict(agent_cfg), log_dir=log_dir, device=device)
     if agent_cfg.class_name == "DistillationRunner":
-        return DistillationRunner(env, to_dict(agent_cfg), log_dir=log_dir, device=agent_cfg.device)
+        return DistillationRunner(env, to_dict(agent_cfg), log_dir=log_dir, device=device)
     raise ValueError(f"Unsupported runner class: {agent_cfg.class_name}")
 
 
diff --git a/source/isaaclab_rl/test/test_rsl_rl_runner.py b/source/isaaclab_rl/test/test_rsl_rl_runner.py
new file mode 100644
index 000000000000..ba01cca82c5c
--- /dev/null
+++ b/source/isaaclab_rl/test/test_rsl_rl_runner.py
@@ -0,0 +1,69 @@
+# Copyright (c) 2022-2026, The Isaac Lab Project Developers (https://github.com/isaac-sim/IsaacLab/blob/main/CONTRIBUTORS.md).
+# All rights reserved.
+#
+# SPDX-License-Identifier: BSD-3-Clause
+
+"""Tests for create_rsl_rl_runner device selection."""
+
+import torch
+from rsl_rl.env import VecEnv
+from tensordict import TensorDict
+
+from isaaclab_rl.rsl_rl import RslRlMLPModelCfg, RslRlOnPolicyRunnerCfg, RslRlPpoAlgorithmCfg
+from isaaclab_rl.rsl_rl.utils import check_rsl_rl_version, create_rsl_rl_runner, handle_deprecated_rsl_rl_cfg
+
+
+class _CpuEnv(VecEnv):
+    """Minimal CPU environment exposing the attributes the runner reads at construction."""
+
+    def __init__(self):
+        self.num_envs = 2
+        self.num_actions = 1
+        self.max_episode_length = 10
+        self.episode_length_buf = torch.zeros(self.num_envs, dtype=torch.long)
+        self.device = "cpu"
+        self.cfg = {}
+
+    def get_observations(self) -> TensorDict:
+        return TensorDict({"policy": torch.zeros(self.num_envs, 3)}, batch_size=[self.num_envs])
+
+    def step(self, actions):
+        raise NotImplementedError
+
+
+def _runner_cfg() -> RslRlOnPolicyRunnerCfg:
+    model = RslRlMLPModelCfg(hidden_dims=[8], activation="elu")
+    cfg = RslRlOnPolicyRunnerCfg(
+        num_steps_per_env=4,
+        max_iterations=1,
+        save_interval=1,
+        experiment_name="test",
+        obs_groups={"actor": ["policy"], "critic": ["policy"]},
+        actor=model.replace(stochastic=True, init_noise_std=1.0),
+        critic=model,
+        algorithm=RslRlPpoAlgorithmCfg(
+            num_learning_epochs=1,
+            num_mini_batches=1,
+            learning_rate=1e-3,
+            schedule="fixed",
+            gamma=0.99,
+            lam=0.95,
+            entropy_coef=0.0,
+            desired_kl=0.01,
+            max_grad_norm=1.0,
+        ),
+    )
+    # the entrypoints adapt the configuration to the installed rsl-rl before building the runner
+    return handle_deprecated_rsl_rl_cfg(cfg, check_rsl_rl_version())
+
+
+def test_runner_falls_back_to_env_device_without_cuda(monkeypatch):
+    """The ``cuda:0`` default must not be used on hosts without CUDA, such as macOS."""
+    monkeypatch.setattr(torch.cuda, "is_available", lambda: False)
+    agent_cfg = _runner_cfg()
+    assert agent_cfg.device == "cuda:0"
+
+    runner = create_rsl_rl_runner(_CpuEnv(), agent_cfg)
+
+    assert runner.device == "cpu"
+    assert all(p.device.type == "cpu" for p in runner.alg.actor.parameters())

From f2cf68f4afa27ace8da51d67560d3dbabb387c2b Mon Sep 17 00:00:00 2001
From: Antoine RICHARD <antoiner@nvidia.com>
Date: Wed, 30 Sep 2026 16:08:01 +0200
Subject: [PATCH 2/9] Warn when the RSL-RL runner falls back to the env device

---
 .../antoiner-fix-rsl-rl-runner-cpu-fallback.rst        |  3 ++-
 source/isaaclab_rl/isaaclab_rl/rsl_rl/utils.py         |  9 +++++++++
 source/isaaclab_rl/test/test_rsl_rl_runner.py          | 10 +++++++---
 3 files changed, 18 insertions(+), 4 deletions(-)

diff --git a/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.rst b/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.rst
index 1ed4157a25ea..7286b5d6ba1f 100644
--- a/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.rst
+++ b/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.rst
@@ -2,4 +2,5 @@ Fixed
 ^^^^^
 
 * Fixed RSL-RL training, playback, and export failing on hosts without CUDA, such as macOS, by
-  building the runner on the environment's device when the agent's default ``cuda:0`` is unavailable.
+  building the runner on the environment's device, with a warning, when the agent's default
+  ``cuda:0`` is unavailable.
diff --git a/source/isaaclab_rl/isaaclab_rl/rsl_rl/utils.py b/source/isaaclab_rl/isaaclab_rl/rsl_rl/utils.py
index 4204f3761ccd..a8b2b1a14d81 100644
--- a/source/isaaclab_rl/isaaclab_rl/rsl_rl/utils.py
+++ b/source/isaaclab_rl/isaaclab_rl/rsl_rl/utils.py
@@ -9,6 +9,7 @@
 
 import importlib.metadata
 import inspect
+import logging
 from dataclasses import MISSING
 from typing import TYPE_CHECKING, Any
 
@@ -31,6 +32,8 @@
 if TYPE_CHECKING:
     from rsl_rl.env import VecEnv
 
+logger = logging.getLogger(__name__)
+
 RSL_RL_MIN_VERSION = "5.0.1"
 """Oldest rsl-rl-lib release supported by the entrypoints."""
 
@@ -74,6 +77,12 @@ def create_rsl_rl_runner(
     """
     device = agent_cfg.device
     if device.startswith("cuda") and not torch.cuda.is_available():
+        logger.warning(
+            "Agent device '%s' is unavailable because CUDA is not available; using the environment device '%s'"
+            " instead.",
+            device,
+            env.device,
+        )
         device = str(env.device)
     if agent_cfg.class_name == "OnPolicyRunner":
         return OnPolicyRunner(env, to_dict(agent_cfg), log_dir=log_dir, device=device)
diff --git a/source/isaaclab_rl/test/test_rsl_rl_runner.py b/source/isaaclab_rl/test/test_rsl_rl_runner.py
index ba01cca82c5c..f6d5b9c3c715 100644
--- a/source/isaaclab_rl/test/test_rsl_rl_runner.py
+++ b/source/isaaclab_rl/test/test_rsl_rl_runner.py
@@ -5,6 +5,8 @@
 
 """Tests for create_rsl_rl_runner device selection."""
 
+import logging
+
 import torch
 from rsl_rl.env import VecEnv
 from tensordict import TensorDict
@@ -57,13 +59,15 @@ def _runner_cfg() -> RslRlOnPolicyRunnerCfg:
     return handle_deprecated_rsl_rl_cfg(cfg, check_rsl_rl_version())
 
 
-def test_runner_falls_back_to_env_device_without_cuda(monkeypatch):
-    """The ``cuda:0`` default must not be used on hosts without CUDA, such as macOS."""
+def test_runner_falls_back_to_env_device_without_cuda(monkeypatch, caplog):
+    """The ``cuda:0`` default must not be used on hosts without CUDA, such as macOS, and the switch is reported."""
     monkeypatch.setattr(torch.cuda, "is_available", lambda: False)
     agent_cfg = _runner_cfg()
     assert agent_cfg.device == "cuda:0"
 
-    runner = create_rsl_rl_runner(_CpuEnv(), agent_cfg)
+    with caplog.at_level(logging.WARNING, logger="isaaclab_rl.rsl_rl.utils"):
+        runner = create_rsl_rl_runner(_CpuEnv(), agent_cfg)
 
     assert runner.device == "cpu"
     assert all(p.device.type == "cpu" for p in runner.alg.actor.parameters())
+    assert any("'cuda:0' is unavailable" in record.getMessage() for record in caplog.records)

From 2284297965141af0e3f0d91f7d6905e146a579e9 Mon Sep 17 00:00:00 2001
From: Antoine RICHARD <antoiner@nvidia.com>
Date: Wed, 30 Sep 2026 18:52:49 +0200
Subject: [PATCH 3/9] Drop the runner device test and docstring note

---
 .../isaaclab_rl/isaaclab_rl/rsl_rl/utils.py   |  3 -
 source/isaaclab_rl/test/test_rsl_rl_runner.py | 73 -------------------
 2 files changed, 76 deletions(-)
 delete mode 100644 source/isaaclab_rl/test/test_rsl_rl_runner.py

diff --git a/source/isaaclab_rl/isaaclab_rl/rsl_rl/utils.py b/source/isaaclab_rl/isaaclab_rl/rsl_rl/utils.py
index a8b2b1a14d81..0f853c68bec7 100644
--- a/source/isaaclab_rl/isaaclab_rl/rsl_rl/utils.py
+++ b/source/isaaclab_rl/isaaclab_rl/rsl_rl/utils.py
@@ -64,9 +64,6 @@ def create_rsl_rl_runner(
 ) -> OnPolicyRunner | DistillationRunner:
     """Instantiate the RSL-RL runner selected by ``agent_cfg.class_name``.
 
-    The runner uses ``agent_cfg.device``, or the environment's device when that is a CUDA device and
-    CUDA is unavailable (e.g. on macOS), since the agent configuration defaults to ``cuda:0``.
-
     Args:
         env: Wrapped environment the runner trains or evaluates on.
         agent_cfg: Runner configuration.
diff --git a/source/isaaclab_rl/test/test_rsl_rl_runner.py b/source/isaaclab_rl/test/test_rsl_rl_runner.py
deleted file mode 100644
index f6d5b9c3c715..000000000000
--- a/source/isaaclab_rl/test/test_rsl_rl_runner.py
+++ /dev/null
@@ -1,73 +0,0 @@
-# Copyright (c) 2022-2026, The Isaac Lab Project Developers (https://github.com/isaac-sim/IsaacLab/blob/main/CONTRIBUTORS.md).
-# All rights reserved.
-#
-# SPDX-License-Identifier: BSD-3-Clause
-
-"""Tests for create_rsl_rl_runner device selection."""
-
-import logging
-
-import torch
-from rsl_rl.env import VecEnv
-from tensordict import TensorDict
-
-from isaaclab_rl.rsl_rl import RslRlMLPModelCfg, RslRlOnPolicyRunnerCfg, RslRlPpoAlgorithmCfg
-from isaaclab_rl.rsl_rl.utils import check_rsl_rl_version, create_rsl_rl_runner, handle_deprecated_rsl_rl_cfg
-
-
-class _CpuEnv(VecEnv):
-    """Minimal CPU environment exposing the attributes the runner reads at construction."""
-
-    def __init__(self):
-        self.num_envs = 2
-        self.num_actions = 1
-        self.max_episode_length = 10
-        self.episode_length_buf = torch.zeros(self.num_envs, dtype=torch.long)
-        self.device = "cpu"
-        self.cfg = {}
-
-    def get_observations(self) -> TensorDict:
-        return TensorDict({"policy": torch.zeros(self.num_envs, 3)}, batch_size=[self.num_envs])
-
-    def step(self, actions):
-        raise NotImplementedError
-
-
-def _runner_cfg() -> RslRlOnPolicyRunnerCfg:
-    model = RslRlMLPModelCfg(hidden_dims=[8], activation="elu")
-    cfg = RslRlOnPolicyRunnerCfg(
-        num_steps_per_env=4,
-        max_iterations=1,
-        save_interval=1,
-        experiment_name="test",
-        obs_groups={"actor": ["policy"], "critic": ["policy"]},
-        actor=model.replace(stochastic=True, init_noise_std=1.0),
-        critic=model,
-        algorithm=RslRlPpoAlgorithmCfg(
-            num_learning_epochs=1,
-            num_mini_batches=1,
-            learning_rate=1e-3,
-            schedule="fixed",
-            gamma=0.99,
-            lam=0.95,
-            entropy_coef=0.0,
-            desired_kl=0.01,
-            max_grad_norm=1.0,
-        ),
-    )
-    # the entrypoints adapt the configuration to the installed rsl-rl before building the runner
-    return handle_deprecated_rsl_rl_cfg(cfg, check_rsl_rl_version())
-
-
-def test_runner_falls_back_to_env_device_without_cuda(monkeypatch, caplog):
-    """The ``cuda:0`` default must not be used on hosts without CUDA, such as macOS, and the switch is reported."""
-    monkeypatch.setattr(torch.cuda, "is_available", lambda: False)
-    agent_cfg = _runner_cfg()
-    assert agent_cfg.device == "cuda:0"
-
-    with caplog.at_level(logging.WARNING, logger="isaaclab_rl.rsl_rl.utils"):
-        runner = create_rsl_rl_runner(_CpuEnv(), agent_cfg)
-
-    assert runner.device == "cpu"
-    assert all(p.device.type == "cpu" for p in runner.alg.actor.parameters())
-    assert any("'cuda:0' is unavailable" in record.getMessage() for record in caplog.records)

From 566ac34fe7ed0d7668aed993e14f929d83892be7 Mon Sep 17 00:00:00 2001
From: Antoine RICHARD <antoiner@nvidia.com>
Date: Thu, 1 Oct 2026 17:51:46 +0200
Subject: [PATCH 4/9] Run RL agents on the resolved simulation device and fall
 back to CPU without CUDA

---
 ...r-fix-rsl-rl-runner-cpu-fallback.changed.rst |  2 ++
 ...ner-fix-rsl-rl-runner-cpu-fallback.fixed.rst |  2 ++
 source/isaaclab/isaaclab/app/sim_launcher.py    |  6 +++++-
 .../rl_games/benchmark_play_rl_games.py         |  2 ++
 .../rl_games/benchmark_train_rl_games.py        |  8 ++++----
 .../backends/rsl_rl/benchmark_play_rsl_rl.py    |  2 ++
 .../backends/rsl_rl/benchmark_train_rsl_rl.py   |  7 ++++---
 .../backends/sb3/benchmark_play_sb3.py          |  2 +-
 .../backends/sb3/benchmark_train_sb3.py         |  2 ++
 ...r-fix-rsl-rl-runner-cpu-fallback.changed.rst |  6 ++++++
 ...ntoiner-fix-rsl-rl-runner-cpu-fallback.major |  0
 .../antoiner-fix-rsl-rl-runner-cpu-fallback.rst |  6 ------
 .../entrypoints/backends/export_rl_games.py     |  2 ++
 .../entrypoints/backends/export_rsl_rl.py       |  2 ++
 .../entrypoints/backends/play_rl_games.py       |  2 ++
 .../entrypoints/backends/play_rsl_rl.py         |  2 ++
 .../entrypoints/backends/play_sb3.py            |  2 +-
 .../entrypoints/backends/train_rl_games.py      |  4 ++--
 .../entrypoints/backends/train_rsl_rl.py        |  3 ++-
 .../entrypoints/backends/train_sb3.py           |  2 ++
 source/isaaclab_rl/isaaclab_rl/rsl_rl/utils.py  | 17 ++---------------
 21 files changed, 47 insertions(+), 34 deletions(-)
 create mode 100644 source/isaaclab/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst
 create mode 100644 source/isaaclab/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.fixed.rst
 create mode 100644 source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst
 create mode 100644 source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.major
 delete mode 100644 source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.rst

diff --git a/source/isaaclab/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst b/source/isaaclab/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst
new file mode 100644
index 000000000000..35d6c2474aa2
--- /dev/null
+++ b/source/isaaclab/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst
@@ -0,0 +1,2 @@
+* Changed the RSL-RL, RL-Games, and Stable-Baselines3 benchmark workflows to run the agent on the device
+  :func:`~isaaclab.app.launch_simulation` resolved for the simulation, matching the training workflows.
diff --git a/source/isaaclab/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.fixed.rst b/source/isaaclab/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.fixed.rst
new file mode 100644
index 000000000000..f84d060eeb84
--- /dev/null
+++ b/source/isaaclab/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.fixed.rst
@@ -0,0 +1,2 @@
+* Fixed runs failing on hosts without CUDA, such as macOS, when the resolved device was a CUDA device.
+  :func:`~isaaclab.app.launch_simulation` now falls back to ``cpu`` with a warning.
diff --git a/source/isaaclab/isaaclab/app/sim_launcher.py b/source/isaaclab/isaaclab/app/sim_launcher.py
index e0d47fdd47ed..cf6dab4811d4 100644
--- a/source/isaaclab/isaaclab/app/sim_launcher.py
+++ b/source/isaaclab/isaaclab/app/sim_launcher.py
@@ -508,11 +508,15 @@ def _resolve_device(sim_cfg, args: dict, launchers: list[SimulationLauncher]) ->
     """Write the run's device to ``sim_cfg.device`` and the ``device`` launcher argument, once.
 
     Starts from the ``device`` launcher argument resolved before launch; a started runtime may refine it
-    (e.g. XR selects the CPU), and a bare ``"cuda"`` is pinned to the physics GPU index.
+    (e.g. XR selects the CPU), a CUDA device falls back to the CPU when CUDA is unavailable (e.g. on macOS),
+    and a bare ``"cuda"`` is pinned to the physics GPU index.
     """
     device = args.get("device")
     for launcher in launchers:
         device = launcher.device or device
+    if device is not None and device.startswith("cuda") and not torch.cuda.is_available():
+        logger.warning("Device '%s' is unavailable because CUDA is not available; using 'cpu' instead.", device)
+        device = "cpu"
     if device == "cuda":
         cuda_device = get_settings_manager().get("/physics/cudaDevice")
         device = f"cuda:{max(0, int(cuda_device) if cuda_device is not None else 0)}"
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_play_rl_games.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_play_rl_games.py
index d7fa1cdbf81f..4467c5ddc73e 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_play_rl_games.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_play_rl_games.py
@@ -186,6 +186,8 @@ def run(argv: list[str]) -> BenchmarkResult:
             )
 
             # Read rl_games device / clipping config.
+            # the agent runs on the device launch_simulation resolved for the simulation
+            agent_cfg["params"]["config"]["device"] = agent_cfg["params"]["config"]["device_name"] = env_cfg.sim.device
             rl_device = agent_cfg["params"]["config"]["device"]
             clip_obs = agent_cfg["params"]["env"].get("clip_observations", math.inf)
             clip_actions = agent_cfg["params"]["env"].get("clip_actions", math.inf)
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_train_rl_games.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_train_rl_games.py
index e004f0ecac4b..f3f334c8e2fa 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_train_rl_games.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_train_rl_games.py
@@ -198,12 +198,12 @@ def run(argv: list[str]) -> BenchmarkResult | None:
             if args_cli.max_iterations is not None:
                 agent_cfg["params"]["config"]["max_epochs"] = args_cli.max_iterations
 
+            # the agent runs on the device launch_simulation resolved for the simulation
+            agent_cfg["params"]["config"]["device"] = agent_cfg["params"]["config"]["device_name"] = env_cfg.sim.device
             if distributed.enabled:
-                # Mirror the regular training entrypoint: the launcher pinned this rank to its own
-                # device, and offsetting the seed by the rank decorrelates exploration across ranks.
+                # Mirror the regular training entrypoint: offsetting the seed by the rank decorrelates
+                # exploration across ranks.
                 agent_cfg["params"]["seed"] += distributed.rank
-                agent_cfg["params"]["config"]["device"] = env_cfg.sim.device
-                agent_cfg["params"]["config"]["device_name"] = env_cfg.sim.device
                 agent_cfg["params"]["config"]["multi_gpu"] = True
             env_cfg.seed = agent_cfg["params"]["seed"]
             horizon_length = agent_cfg["params"]["config"].get("horizon_length", 16)
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_play_rsl_rl.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_play_rsl_rl.py
index 78a83cf463da..7fde3dabf711 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_play_rsl_rl.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_play_rsl_rl.py
@@ -145,6 +145,8 @@ def run(argv: list[str]) -> BenchmarkResult:
 
             installed_rsl_rl = metadata.version("rsl-rl-lib")
             agent_cfg = handle_deprecated_rsl_rl_cfg(agent_cfg, installed_rsl_rl)
+            # the agent runs on the device launch_simulation resolved for the simulation
+            agent_cfg.device = env_cfg.sim.device
 
             log_root_path = os.path.abspath(os.path.join("logs", "rsl_rl", agent_cfg.experiment_name))
             if args.checkpoint in common.CHECKPOINT_SELECTORS:
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_train_rsl_rl.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_train_rsl_rl.py
index 876d54c28368..760f8e3a7b5f 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_train_rsl_rl.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_train_rsl_rl.py
@@ -191,10 +191,11 @@ def run(argv: list[str]) -> BenchmarkResult | None:
             agent_cfg = handle_deprecated_rsl_rl_cfg(agent_cfg, installed_rsl_rl)
             env_cfg.seed = agent_cfg.seed
 
+            # the agent runs on the device launch_simulation resolved for the simulation
+            agent_cfg.device = env_cfg.sim.device
             if distributed.enabled:
-                # Mirror the regular training entrypoint: the launcher pinned this rank to its own
-                # device, and offsetting the seed by the rank decorrelates exploration across ranks.
-                agent_cfg.device = env_cfg.sim.device
+                # Mirror the regular training entrypoint: offsetting the seed by the rank decorrelates
+                # exploration across ranks.
                 agent_cfg.seed += distributed.rank
                 env_cfg.seed = agent_cfg.seed
             reported_num_envs, _ = distributed.global_work(env_cfg.scene.num_envs, agent_cfg.num_steps_per_env)
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_play_sb3.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_play_sb3.py
index ecea49d98402..bbb6de5afd25 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_play_sb3.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_play_sb3.py
@@ -213,7 +213,7 @@ def run(argv: list[str]) -> BenchmarkResult:
                 )
 
             # Load the trained policy.
-            agent = PPO.load(resume_path, env, print_system_info=True)
+            agent = PPO.load(resume_path, env, device=env_cfg.sim.device, print_system_info=True)
 
             def policy(obs):
                 """Map an observation batch to a deterministic action batch via the sb3 agent.
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_train_sb3.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_train_sb3.py
index dabc72fc394b..4065ad073996 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_train_sb3.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_train_sb3.py
@@ -239,6 +239,8 @@ def run(argv: list[str]) -> BenchmarkResult:
 
             agent_cfg["seed"] = args_cli.seed if args_cli.seed is not None else agent_cfg.get("seed", 0)
             env_cfg.seed = agent_cfg["seed"]
+            # the agent runs on the device launch_simulation resolved for the simulation
+            agent_cfg["device"] = env_cfg.sim.device
 
             # Convert the iteration override to SB3 total timesteps.
             n_steps_cfg = agent_cfg.get("n_steps", 2048)
diff --git a/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst b/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst
new file mode 100644
index 000000000000..150adf6f6eef
--- /dev/null
+++ b/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst
@@ -0,0 +1,6 @@
+* **Breaking:** Changed the RSL-RL, RL-Games, and Stable-Baselines3 train, play, and export workflows to run the
+  agent on the device :func:`~isaaclab.app.launch_simulation` resolved for the simulation, as skrl and TorchRL
+  already did. Outside distributed runs they used the agent configuration's device, ``cuda:0`` by default, so
+  ``--device cpu`` trained on ``cuda:0``, ``--device cuda:1`` trained on GPU 0, and hosts without CUDA, such as
+  macOS, failed. Pass ``--device`` to choose where both the simulation and the agent run; the agent
+  configuration's ``device`` is now overwritten by these workflows.
diff --git a/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.major b/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.major
new file mode 100644
index 000000000000..e69de29bb2d1
diff --git a/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.rst b/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.rst
deleted file mode 100644
index 7286b5d6ba1f..000000000000
--- a/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.rst
+++ /dev/null
@@ -1,6 +0,0 @@
-Fixed
-^^^^^
-
-* Fixed RSL-RL training, playback, and export failing on hosts without CUDA, such as macOS, by
-  building the runner on the environment's device, with a warning, when the agent's default
-  ``cuda:0`` is unavailable.
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rl_games.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rl_games.py
index 48ac9f627876..f710f7343b87 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rl_games.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rl_games.py
@@ -103,6 +103,8 @@ def export_rl_games_agent(args_cli: argparse.Namespace, env_cfg: Any, agent_cfg:
     from leapp import annotate
 
     params = agent_cfg["params"]
+    # the agent runs on the device launch_simulation resolved for the simulation
+    params["config"]["device"] = params["config"]["device_name"] = env_cfg.sim.device
     env_cfg.scene.num_envs = 1
     env_cfg.seed = params["seed"]
 
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rsl_rl.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rsl_rl.py
index c97db09d2d2a..c6cb9d40abe4 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rsl_rl.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rsl_rl.py
@@ -140,6 +140,8 @@ def export_rsl_rl_agent(args_cli: argparse.Namespace, env_cfg: Any, agent_cfg: R
     if args_cli.experiment_name is not None:
         agent_cfg.experiment_name = args_cli.experiment_name
     agent_cfg = handle_deprecated_rsl_rl_cfg(agent_cfg, installed_version)
+    # the agent runs on the device launch_simulation resolved for the simulation
+    agent_cfg.device = env_cfg.sim.device
 
     env_cfg.scene.num_envs = 1
     # certain randomizations occur in the environment initialization so we set the seed here
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rl_games.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rl_games.py
index f1b33a7ab270..f7fe9c566f95 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rl_games.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rl_games.py
@@ -110,6 +110,8 @@ def run(argv: list[str]) -> None:
             show_run_summary(screen, args_cli, env_cfg, library="rl_games", action="play")
             apply_env_overrides(args_cli, env_cfg)
             params = agent_cfg["params"]
+            # the agent runs on the device launch_simulation resolved for the simulation
+            params["config"]["device"] = params["config"]["device_name"] = env_cfg.sim.device
             args_cli.seed = resolve_seed(args_cli.seed)
             if args_cli.seed is not None:
                 params["seed"] = args_cli.seed
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rsl_rl.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rsl_rl.py
index efeca2b79e9b..2ccb79760c70 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rsl_rl.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rsl_rl.py
@@ -111,6 +111,8 @@ def run(argv: list[str]) -> None:
             show_run_summary(screen, args_cli, env_cfg, library="rsl_rl", action="play")
             agent_cfg = cli_args.update_rsl_rl_cfg(agent_cfg, args_cli)
             agent_cfg = handle_deprecated_rsl_rl_cfg(agent_cfg, installed_version)
+            # the agent runs on the device launch_simulation resolved for the simulation
+            agent_cfg.device = env_cfg.sim.device
             apply_env_overrides(args_cli, env_cfg)
             # certain randomizations occur in the environment initialization so we set the seed here
             env_cfg.seed = agent_cfg.seed
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_sb3.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_sb3.py
index aec7d0440aab..e07a68975f85 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_sb3.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_sb3.py
@@ -144,7 +144,7 @@ def run(argv: list[str]) -> None:
                 )
 
             print(f"Loading checkpoint from: {checkpoint_path}")
-            agent = PPO.load(checkpoint_path, env, print_system_info=True)
+            agent = PPO.load(checkpoint_path, env, device=env_cfg.sim.device, print_system_info=True)
             # configure_seed must run after PPO.load so torch determinism does not disturb SB3's initialization
             if args_cli.deterministic:
                 configure_seed(env_cfg.seed, torch_deterministic=True)
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rl_games.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rl_games.py
index c0263b6748bc..00d34bd4afa5 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rl_games.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rl_games.py
@@ -118,11 +118,11 @@ def run(argv: list[str]) -> None:
                 params["seed"] = args_cli.seed
             if args_cli.max_iterations is not None:
                 config["max_epochs"] = args_cli.max_iterations
+            # the agent runs on the device launch_simulation resolved for the simulation
+            config["device"] = config["device_name"] = env_cfg.sim.device
             rank = int(os.getenv("RANK", "0")) if args_cli.distributed else None
             if rank is not None:
                 params["seed"] += rank
-                config["device"] = env_cfg.sim.device
-                config["device_name"] = env_cfg.sim.device
                 config["multi_gpu"] = True
             env_cfg.seed = params["seed"]
 
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rsl_rl.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rsl_rl.py
index beeab8d41679..6a0ec3acf174 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rsl_rl.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rsl_rl.py
@@ -128,9 +128,10 @@ def _run(args_cli: argparse.Namespace) -> None:
             apply_env_overrides(args_cli, env_cfg)
             if args_cli.max_iterations is not None:
                 agent_cfg.max_iterations = args_cli.max_iterations
+            # the agent runs on the device launch_simulation resolved for the simulation
+            agent_cfg.device = env_cfg.sim.device
             rank = int(os.getenv("RANK", "0")) if args_cli.distributed else None
             if rank is not None:
-                agent_cfg.device = env_cfg.sim.device
                 agent_cfg.seed += rank
             env_cfg.seed = agent_cfg.seed
 
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_sb3.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_sb3.py
index 92daf0bff64f..491b667f9bd2 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_sb3.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_sb3.py
@@ -118,6 +118,8 @@ def run(argv: list[str]) -> None:
             if args_cli.max_iterations is not None:
                 agent_cfg["n_timesteps"] = args_cli.max_iterations * agent_cfg["n_steps"] * env_cfg.scene.num_envs
             env_cfg.seed = agent_cfg["seed"]
+            # the agent runs on the device launch_simulation resolved for the simulation
+            agent_cfg["device"] = env_cfg.sim.device
 
             log_root_path = os.path.abspath(os.path.join("logs", "sb3", args_cli.task))
             print(f"[INFO] Logging experiment in directory: {log_root_path}")
diff --git a/source/isaaclab_rl/isaaclab_rl/rsl_rl/utils.py b/source/isaaclab_rl/isaaclab_rl/rsl_rl/utils.py
index 0f853c68bec7..8505e2b4a7a4 100644
--- a/source/isaaclab_rl/isaaclab_rl/rsl_rl/utils.py
+++ b/source/isaaclab_rl/isaaclab_rl/rsl_rl/utils.py
@@ -9,11 +9,9 @@
 
 import importlib.metadata
 import inspect
-import logging
 from dataclasses import MISSING
 from typing import TYPE_CHECKING, Any
 
-import torch
 from packaging import version
 from rsl_rl.runners import DistillationRunner, OnPolicyRunner
 
@@ -32,8 +30,6 @@
 if TYPE_CHECKING:
     from rsl_rl.env import VecEnv
 
-logger = logging.getLogger(__name__)
-
 RSL_RL_MIN_VERSION = "5.0.1"
 """Oldest rsl-rl-lib release supported by the entrypoints."""
 
@@ -72,19 +68,10 @@ def create_rsl_rl_runner(
     Raises:
         ValueError: If the configured runner class is not supported.
     """
-    device = agent_cfg.device
-    if device.startswith("cuda") and not torch.cuda.is_available():
-        logger.warning(
-            "Agent device '%s' is unavailable because CUDA is not available; using the environment device '%s'"
-            " instead.",
-            device,
-            env.device,
-        )
-        device = str(env.device)
     if agent_cfg.class_name == "OnPolicyRunner":
-        return OnPolicyRunner(env, to_dict(agent_cfg), log_dir=log_dir, device=device)
+        return OnPolicyRunner(env, to_dict(agent_cfg), log_dir=log_dir, device=agent_cfg.device)
     if agent_cfg.class_name == "DistillationRunner":
-        return DistillationRunner(env, to_dict(agent_cfg), log_dir=log_dir, device=device)
+        return DistillationRunner(env, to_dict(agent_cfg), log_dir=log_dir, device=agent_cfg.device)
     raise ValueError(f"Unsupported runner class: {agent_cfg.class_name}")
 
 

From 0c81e675ec59fe8a76d642f9e3b0566a4804231a Mon Sep 17 00:00:00 2001
From: Antoine RICHARD <antoiner@nvidia.com>
Date: Mon, 5 Oct 2026 09:27:06 +0200
Subject: [PATCH 5/9] Set the agent device in each RL library's shared config
 setup and raise without CUDA

---
 ...fix-rsl-rl-runner-cpu-fallback.changed.rst |  4 +-
 ...r-fix-rsl-rl-runner-cpu-fallback.fixed.rst |  4 +-
 source/isaaclab/isaaclab/app/sim_launcher.py  |  9 ++--
 .../rl_games/benchmark_play_rl_games.py       |  6 +--
 .../rl_games/benchmark_train_rl_games.py      | 17 +-------
 .../backends/rsl_rl/benchmark_play_rsl_rl.py  |  7 +---
 .../backends/rsl_rl/benchmark_train_rsl_rl.py |  2 -
 .../backends/sb3/benchmark_play_sb3.py        |  3 +-
 .../backends/sb3/benchmark_train_sb3.py       |  5 +--
 ...fix-rsl-rl-runner-cpu-fallback.changed.rst |  8 ++--
 .../entrypoints/backends/cli_args_rl_games.py | 41 +++++++++++++++++++
 .../entrypoints/backends/cli_args_rsl_rl.py   | 14 ++++---
 .../entrypoints/backends/cli_args_sb3.py      | 32 +++++++++++++++
 .../entrypoints/backends/export_rl_games.py   |  4 +-
 .../entrypoints/backends/export_rsl_rl.py     | 12 ++----
 .../entrypoints/backends/play_rl_games.py     |  8 +---
 .../entrypoints/backends/play_rsl_rl.py       |  2 -
 .../entrypoints/backends/play_sb3.py          |  6 +--
 .../entrypoints/backends/train_rl_games.py    | 13 +-----
 .../entrypoints/backends/train_rsl_rl.py      |  2 -
 .../entrypoints/backends/train_sb3.py         |  8 +---
 .../test/test_entrypoints_common.py           |  4 +-
 22 files changed, 121 insertions(+), 90 deletions(-)
 create mode 100644 source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rl_games.py
 create mode 100644 source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_sb3.py

diff --git a/source/isaaclab/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst b/source/isaaclab/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst
index 35d6c2474aa2..a8fe1c93da74 100644
--- a/source/isaaclab/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst
+++ b/source/isaaclab/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst
@@ -1,2 +1,2 @@
-* Changed the RSL-RL, RL-Games, and Stable-Baselines3 benchmark workflows to run the agent on the device
-  :func:`~isaaclab.app.launch_simulation` resolved for the simulation, matching the training workflows.
+* Changed the RSL-RL, RL-Games, and Stable-Baselines3 benchmark workflows to set up their agent configuration with
+  the same shared overrides as the training and play workflows, including the device the simulation runs on.
diff --git a/source/isaaclab/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.fixed.rst b/source/isaaclab/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.fixed.rst
index f84d060eeb84..82d1534f9b60 100644
--- a/source/isaaclab/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.fixed.rst
+++ b/source/isaaclab/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.fixed.rst
@@ -1,2 +1,2 @@
-* Fixed runs failing on hosts without CUDA, such as macOS, when the resolved device was a CUDA device.
-  :func:`~isaaclab.app.launch_simulation` now falls back to ``cpu`` with a warning.
+* Fixed runs on hosts without CUDA, such as macOS, failing inside PyTorch when the resolved device was a CUDA
+  device. :func:`~isaaclab.app.launch_simulation` now raises a clear error that asks for ``--device cpu``.
diff --git a/source/isaaclab/isaaclab/app/sim_launcher.py b/source/isaaclab/isaaclab/app/sim_launcher.py
index da00f5fc2380..64bcd84c8249 100644
--- a/source/isaaclab/isaaclab/app/sim_launcher.py
+++ b/source/isaaclab/isaaclab/app/sim_launcher.py
@@ -503,15 +503,16 @@ def _resolve_device(sim_cfg, args: dict, launchers: list[SimulationLauncher]) ->
     """Write the run's device to ``sim_cfg.device`` and the ``device`` launcher argument, once.
 
     Starts from the ``device`` launcher argument resolved before launch; a started runtime may refine it
-    (e.g. XR selects the CPU), a CUDA device falls back to the CPU when CUDA is unavailable (e.g. on macOS),
-    and a bare ``"cuda"`` is pinned to the physics GPU index.
+    (e.g. XR selects the CPU), and a bare ``"cuda"`` is pinned to the physics GPU index.
+
+    Raises:
+        RuntimeError: When the device is a CUDA device and CUDA is unavailable (e.g. on macOS).
     """
     device = args.get("device")
     for launcher in launchers:
         device = launcher.device or device
     if device is not None and device.startswith("cuda") and not torch.cuda.is_available():
-        logger.warning("Device '%s' is unavailable because CUDA is not available; using 'cpu' instead.", device)
-        device = "cpu"
+        raise RuntimeError(f"Device '{device}' requires CUDA, which is not available; pass --device cpu.")
     if device == "cuda":
         cuda_device = get_settings_manager().get("/physics/cudaDevice")
         device = f"cuda:{max(0, int(cuda_device) if cuda_device is not None else 0)}"
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_play_rl_games.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_play_rl_games.py
index 3fd9b690cf10..e928c9ec9fde 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_play_rl_games.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_play_rl_games.py
@@ -113,6 +113,7 @@ def run(argv: list[str]) -> BenchmarkResult:
     from isaaclab.benchmark import BaseIsaacLabBenchmark, BenchmarkMonitor, BenchmarkResult, builders, capture, stepping
     from isaaclab.benchmark.schema import StartupTime
 
+    from isaaclab_rl.entrypoints.backends import cli_args_rl_games as cli_args
     from isaaclab_rl.rl_games import RlGamesGpuEnv, RlGamesVecEnvWrapper
 
     # Importing the task packages registers their gym environments so the
@@ -139,8 +140,7 @@ def run(argv: list[str]) -> BenchmarkResult:
 
             if args_cli.num_envs is not None:
                 env_cfg.scene.num_envs = args_cli.num_envs
-            if args_cli.seed is not None:
-                agent_cfg["params"]["seed"] = args_cli.seed
+            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli)
             env_cfg.seed = agent_cfg["params"]["seed"]
 
             config_name = agent_cfg["params"]["config"]["name"]
@@ -185,8 +185,6 @@ def run(argv: list[str]) -> BenchmarkResult:
             )
 
             # Read rl_games device / clipping config.
-            # the agent runs on the device launch_simulation resolved for the simulation
-            agent_cfg["params"]["config"]["device"] = agent_cfg["params"]["config"]["device_name"] = env_cfg.sim.device
             rl_device = agent_cfg["params"]["config"]["device"]
             clip_obs = agent_cfg["params"]["env"].get("clip_observations", math.inf)
             clip_actions = agent_cfg["params"]["env"].get("clip_actions", math.inf)
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_train_rl_games.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_train_rl_games.py
index 8bd9f6289560..1582cfd45887 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_train_rl_games.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_train_rl_games.py
@@ -129,7 +129,6 @@ def run(argv: list[str]) -> BenchmarkResult | None:
     import contextlib
     import math
     import os
-    import random
     from datetime import datetime
 
     from rl_games.common import env_configurations, vecenv
@@ -150,6 +149,7 @@ def run(argv: list[str]) -> BenchmarkResult | None:
     from isaaclab.benchmark.metrics import RL_LIBRARY_DESCRIPTORS, parse_tf_logs
     from isaaclab.benchmark.schema import StartupTime
 
+    from isaaclab_rl.entrypoints.backends import cli_args_rl_games as cli_args
     from isaaclab_rl.rl_games import RlGamesGpuEnv, RlGamesVecEnvWrapper
 
     import isaaclab_tasks  # noqa: F401
@@ -194,20 +194,7 @@ def run(argv: list[str]) -> BenchmarkResult | None:
 
             apply_env_overrides(args_cli, env_cfg)
 
-            if args_cli.seed == -1:
-                args_cli.seed = random.randint(0, 10000)
-            agent_cfg["params"]["seed"] = args_cli.seed if args_cli.seed is not None else agent_cfg["params"]["seed"]
-
-            if args_cli.max_iterations is not None:
-                agent_cfg["params"]["config"]["max_epochs"] = args_cli.max_iterations
-
-            # the agent runs on the device launch_simulation resolved for the simulation
-            agent_cfg["params"]["config"]["device"] = agent_cfg["params"]["config"]["device_name"] = env_cfg.sim.device
-            if distributed.enabled:
-                # Mirror the regular training entrypoint: offsetting the seed by the rank decorrelates
-                # exploration across ranks.
-                agent_cfg["params"]["seed"] += distributed.rank
-                agent_cfg["params"]["config"]["multi_gpu"] = True
+            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli)
             env_cfg.seed = agent_cfg["params"]["seed"]
             horizon_length = agent_cfg["params"]["config"].get("horizon_length", 16)
             reported_num_envs, _ = distributed.global_work(env_cfg.scene.num_envs, horizon_length)
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_play_rsl_rl.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_play_rsl_rl.py
index 2b001635805d..c5a5a2b53a25 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_play_rsl_rl.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_play_rsl_rl.py
@@ -111,6 +111,7 @@ def run(argv: list[str]) -> BenchmarkResult:
     from isaaclab.benchmark import BaseIsaacLabBenchmark, BenchmarkMonitor, BenchmarkResult, builders, capture, stepping
     from isaaclab.benchmark.schema import StartupTime
 
+    from isaaclab_rl.entrypoints.backends import cli_args_rsl_rl as cli_args
     from isaaclab_rl.rsl_rl import RslRlVecEnvWrapper
 
     # Importing the task packages registers their gym environments so the
@@ -137,13 +138,9 @@ def run(argv: list[str]) -> BenchmarkResult:
 
             if args.num_envs is not None:
                 env_cfg.scene.num_envs = args.num_envs
-            if args.seed is not None:
-                agent_cfg.seed = args.seed
+            agent_cfg = cli_args.update_rsl_rl_cfg(agent_cfg, args)
             env_cfg.seed = agent_cfg.seed
 
-            # the agent runs on the device launch_simulation resolved for the simulation
-            agent_cfg.device = env_cfg.sim.device
-
             log_root_path = os.path.abspath(os.path.join("logs", "rsl_rl", agent_cfg.experiment_name))
             if args.checkpoint in common.CHECKPOINT_SELECTORS:
                 resume_path = common.resolve_checkpoint_selector(
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_train_rsl_rl.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_train_rsl_rl.py
index 59a4d884c64f..dfd5f2af1e61 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_train_rsl_rl.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_train_rsl_rl.py
@@ -191,8 +191,6 @@ def run(argv: list[str]) -> BenchmarkResult | None:
             )
             env_cfg.seed = agent_cfg.seed
 
-            # the agent runs on the device launch_simulation resolved for the simulation
-            agent_cfg.device = env_cfg.sim.device
             if distributed.enabled:
                 # Mirror the regular training entrypoint: offsetting the seed by the rank decorrelates
                 # exploration across ranks.
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_play_sb3.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_play_sb3.py
index 4ffd3550e233..8d4671cc7415 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_play_sb3.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_play_sb3.py
@@ -116,6 +116,7 @@ def run(argv: list[str]) -> BenchmarkResult:
     from isaaclab.benchmark import BaseIsaacLabBenchmark, BenchmarkMonitor, BenchmarkResult, builders, capture, stepping
     from isaaclab.benchmark.schema import StartupTime
 
+    from isaaclab_rl.entrypoints.backends import cli_args_sb3 as cli_args
     from isaaclab_rl.sb3 import Sb3VecEnvWrapper, process_sb3_cfg
 
     # Importing the task packages registers their gym environments so the
@@ -142,7 +143,7 @@ def run(argv: list[str]) -> BenchmarkResult:
 
             if args_cli.num_envs is not None:
                 env_cfg.scene.num_envs = args_cli.num_envs
-            agent_cfg["seed"] = args_cli.seed if args_cli.seed is not None else agent_cfg.get("seed", 0)
+            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli)
             env_cfg.seed = agent_cfg["seed"]
 
             log_root_path = os.path.abspath(os.path.join("logs", "sb3", args_cli.task))
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_train_sb3.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_train_sb3.py
index ea59d7819b2a..7b23c0cfb17f 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_train_sb3.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_train_sb3.py
@@ -204,6 +204,7 @@ def run(argv: list[str]) -> BenchmarkResult:
     from isaaclab.benchmark.metrics import RL_LIBRARY_DESCRIPTORS, parse_tf_logs
     from isaaclab.benchmark.schema import StartupTime
 
+    from isaaclab_rl.entrypoints.backends import cli_args_sb3 as cli_args
     from isaaclab_rl.sb3 import Sb3VecEnvWrapper, process_sb3_cfg
 
     import isaaclab_tasks  # noqa: F401
@@ -240,10 +241,8 @@ def run(argv: list[str]) -> BenchmarkResult:
 
             apply_env_overrides(args_cli, env_cfg)
 
-            agent_cfg["seed"] = args_cli.seed if args_cli.seed is not None else agent_cfg.get("seed", 0)
+            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli)
             env_cfg.seed = agent_cfg["seed"]
-            # the agent runs on the device launch_simulation resolved for the simulation
-            agent_cfg["device"] = env_cfg.sim.device
 
             # Convert the iteration override to SB3 total timesteps.
             n_steps_cfg = agent_cfg.get("n_steps", 2048)
diff --git a/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst b/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst
index 150adf6f6eef..dec9ccf74c4c 100644
--- a/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst
+++ b/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst
@@ -1,6 +1,8 @@
 * **Breaking:** Changed the RSL-RL, RL-Games, and Stable-Baselines3 train, play, and export workflows to run the
   agent on the device :func:`~isaaclab.app.launch_simulation` resolved for the simulation, as skrl and TorchRL
   already did. Outside distributed runs they used the agent configuration's device, ``cuda:0`` by default, so
-  ``--device cpu`` trained on ``cuda:0``, ``--device cuda:1`` trained on GPU 0, and hosts without CUDA, such as
-  macOS, failed. Pass ``--device`` to choose where both the simulation and the agent run; the agent
-  configuration's ``device`` is now overwritten by these workflows.
+  ``--device cpu`` trained on ``cuda:0`` and ``--device cuda:1`` trained on GPU 0. Pass ``--device`` to choose
+  where both the simulation and the agent run; the agent configuration's ``device`` is now overwritten.
+* Changed each library's agent configuration overrides to live in one shared function that every train, play,
+  export, and benchmark workflow calls: ``update_rsl_rl_cfg``, and the new ``update_rl_games_cfg`` and
+  ``update_sb3_cfg``.
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rl_games.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rl_games.py
new file mode 100644
index 000000000000..a85d04135d32
--- /dev/null
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rl_games.py
@@ -0,0 +1,41 @@
+# Copyright (c) 2022-2026, The Isaac Lab Project Developers (https://github.com/isaac-sim/IsaacLab/blob/main/CONTRIBUTORS.md).
+# All rights reserved.
+#
+# SPDX-License-Identifier: BSD-3-Clause
+
+"""Agent configuration overrides shared by the RL-Games entrypoints."""
+
+from __future__ import annotations
+
+import argparse
+import os
+
+from ..common import resolve_seed
+
+
+def update_rl_games_cfg(agent_cfg: dict, args_cli: argparse.Namespace) -> dict:
+    """Override an RL-Games agent configuration with the command-line arguments.
+
+    Call it inside :func:`~isaaclab.app.launch_simulation`, which resolves ``args_cli.device``: the agent runs on
+    the simulation's device.
+
+    Args:
+        agent_cfg: The configuration for RL-Games agent.
+        args_cli: The command line arguments.
+
+    Returns:
+        The updated RL-Games agent configuration.
+    """
+    params = agent_cfg["params"]
+    config = params["config"]
+    config["device"] = config["device_name"] = args_cli.device
+    if getattr(args_cli, "seed", None) is not None:
+        args_cli.seed = resolve_seed(args_cli.seed)
+        params["seed"] = args_cli.seed
+    if getattr(args_cli, "max_iterations", None) is not None:
+        config["max_epochs"] = args_cli.max_iterations
+    if getattr(args_cli, "distributed", False):
+        # offsetting the seed by the rank decorrelates exploration across ranks
+        params["seed"] += int(os.getenv("RANK", "0"))
+        config["multi_gpu"] = True
+    return agent_cfg
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rsl_rl.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rsl_rl.py
index 0c953c4b44ca..484858cff748 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rsl_rl.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rsl_rl.py
@@ -87,6 +87,9 @@ def parse_rsl_rl_cfg(task_name: str, args_cli: argparse.Namespace) -> RslRlBaseR
 def update_rsl_rl_cfg(agent_cfg: RslRlBaseRunnerCfg, args_cli: argparse.Namespace) -> RslRlBaseRunnerCfg:
     """Override an RSL-RL agent configuration with the command-line arguments.
 
+    Call it inside :func:`~isaaclab.app.launch_simulation`, which resolves ``args_cli.device``: the agent runs on
+    the simulation's device.
+
     Args:
         agent_cfg: The configuration for RSL-RL agent.
         args_cli: The command line arguments.
@@ -94,18 +97,19 @@ def update_rsl_rl_cfg(agent_cfg: RslRlBaseRunnerCfg, args_cli: argparse.Namespac
     Returns:
         The updated RSL-RL agent configuration.
     """
+    agent_cfg.device = args_cli.device
     if getattr(args_cli, "seed", None) is not None:
         args_cli.seed = resolve_seed(args_cli.seed)
         agent_cfg.seed = args_cli.seed
-    if args_cli.checkpoint is not None:
+    if getattr(args_cli, "checkpoint", None) is not None:
         agent_cfg.load_checkpoint = args_cli.checkpoint
-    if args_cli.experiment_name is not None:
+    if getattr(args_cli, "experiment_name", None) is not None:
         agent_cfg.experiment_name = args_cli.experiment_name
-    if args_cli.run_name is not None:
+    if getattr(args_cli, "run_name", None) is not None:
         agent_cfg.run_name = args_cli.run_name
-    if args_cli.logger is not None:
+    if getattr(args_cli, "logger", None) is not None:
         agent_cfg.logger = args_cli.logger
-    if agent_cfg.logger in {"wandb", "neptune"} and args_cli.log_project_name:
+    if agent_cfg.logger in {"wandb", "neptune"} and getattr(args_cli, "log_project_name", None):
         agent_cfg.wandb_project = args_cli.log_project_name
         agent_cfg.neptune_project = args_cli.log_project_name
     return agent_cfg
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_sb3.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_sb3.py
new file mode 100644
index 000000000000..71f3d2cb8cf0
--- /dev/null
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_sb3.py
@@ -0,0 +1,32 @@
+# Copyright (c) 2022-2026, The Isaac Lab Project Developers (https://github.com/isaac-sim/IsaacLab/blob/main/CONTRIBUTORS.md).
+# All rights reserved.
+#
+# SPDX-License-Identifier: BSD-3-Clause
+
+"""Agent configuration overrides shared by the Stable-Baselines3 entrypoints."""
+
+from __future__ import annotations
+
+import argparse
+
+from ..common import resolve_seed
+
+
+def update_sb3_cfg(agent_cfg: dict, args_cli: argparse.Namespace) -> dict:
+    """Override a Stable-Baselines3 agent configuration with the command-line arguments.
+
+    Call it inside :func:`~isaaclab.app.launch_simulation`, which resolves ``args_cli.device``: the agent runs on
+    the simulation's device. Loading a checkpoint still needs that device passed to ``PPO.load``.
+
+    Args:
+        agent_cfg: The configuration for Stable-Baselines3 agent.
+        args_cli: The command line arguments.
+
+    Returns:
+        The updated Stable-Baselines3 agent configuration.
+    """
+    agent_cfg["device"] = args_cli.device
+    if getattr(args_cli, "seed", None) is not None:
+        args_cli.seed = resolve_seed(args_cli.seed)
+        agent_cfg["seed"] = args_cli.seed
+    return agent_cfg
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rl_games.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rl_games.py
index e1cbb305ee5b..7601ceb1be0f 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rl_games.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rl_games.py
@@ -32,6 +32,7 @@
 
 from ...rl_games import RlGamesVecEnvWrapper, register_rl_games_env
 from ..common import resolve_published_checkpoint
+from . import cli_args_rl_games as cli_args
 from .export_common import (
     add_common_export_args,
     finalize_export_args,
@@ -105,9 +106,8 @@ def export_rl_games_agent(args_cli: argparse.Namespace, env_cfg: Any, agent_cfg:
     # initialized the selected backend
     from leapp import annotate
 
+    agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli)
     params = agent_cfg["params"]
-    # the agent runs on the device launch_simulation resolved for the simulation
-    params["config"]["device"] = params["config"]["device_name"] = env_cfg.sim.device
     env_cfg.scene.num_envs = 1
     env_cfg.seed = params["seed"]
 
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rsl_rl.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rsl_rl.py
index 17ae66f527f4..e1fcfd329bb7 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rsl_rl.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rsl_rl.py
@@ -30,7 +30,8 @@
 import isaaclab_tasks  # noqa: F401
 
 from ...rsl_rl import RslRlBaseRunnerCfg, RslRlVecEnvWrapper
-from ..common import resolve_published_checkpoint, resolve_seed
+from ..common import resolve_published_checkpoint
+from . import cli_args_rsl_rl as cli_args
 from .export_common import (
     add_common_export_args,
     finalize_export_args,
@@ -131,14 +132,7 @@ def export_rsl_rl_agent(args_cli: argparse.Namespace, env_cfg: Any, agent_cfg: R
     # the LEAPP runtime loads simulation modules, so import it only after the launch
     from leapp import annotate
 
-    if args_cli.seed is not None:
-        agent_cfg.seed = resolve_seed(args_cli.seed)
-    if args_cli.checkpoint is not None:
-        agent_cfg.load_checkpoint = args_cli.checkpoint
-    if args_cli.experiment_name is not None:
-        agent_cfg.experiment_name = args_cli.experiment_name
-    # the agent runs on the device launch_simulation resolved for the simulation
-    agent_cfg.device = env_cfg.sim.device
+    agent_cfg = cli_args.update_rsl_rl_cfg(agent_cfg, args_cli)
 
     env_cfg.scene.num_envs = 1
     # certain randomizations occur in the environment initialization so we set the seed here
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rl_games.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rl_games.py
index 3ca493db61e7..b4376e0356e1 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rl_games.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rl_games.py
@@ -38,12 +38,12 @@
     pre_launch_video_config,
     resolve_checkpoint_selector,
     resolve_published_checkpoint,
-    resolve_seed,
     run_playback,
     set_hydra_args,
     show_run_summary,
     startup_screen,
 )
+from . import cli_args_rl_games as cli_args
 
 logger = logging.getLogger(__name__)
 
@@ -112,12 +112,8 @@ def run(argv: list[str]) -> None:
         with launch_simulation(env_cfg, args_cli), contextlib.ExitStack() as cleanup:
             show_run_summary(screen, args_cli, env_cfg, library="rl_games", action="play")
             apply_env_overrides(args_cli, env_cfg)
+            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli)
             params = agent_cfg["params"]
-            # the agent runs on the device launch_simulation resolved for the simulation
-            params["config"]["device"] = params["config"]["device_name"] = env_cfg.sim.device
-            args_cli.seed = resolve_seed(args_cli.seed)
-            if args_cli.seed is not None:
-                params["seed"] = args_cli.seed
             env_cfg.seed = params["seed"]
 
             log_root_path = os.path.abspath(os.path.join("logs", "rl_games", params["config"]["name"]))
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rsl_rl.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rsl_rl.py
index 5797e367d2ca..46c32aceef74 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rsl_rl.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rsl_rl.py
@@ -108,8 +108,6 @@ def run(argv: list[str]) -> None:
         with launch_simulation(env_cfg, args_cli), contextlib.ExitStack() as cleanup:
             show_run_summary(screen, args_cli, env_cfg, library="rsl_rl", action="play")
             agent_cfg = cli_args.update_rsl_rl_cfg(agent_cfg, args_cli)
-            # the agent runs on the device launch_simulation resolved for the simulation
-            agent_cfg.device = env_cfg.sim.device
             apply_env_overrides(args_cli, env_cfg)
             # certain randomizations occur in the environment initialization so we set the seed here
             env_cfg.seed = agent_cfg.seed
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_sb3.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_sb3.py
index ef096a6b0da6..3371ffd8e936 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_sb3.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_sb3.py
@@ -36,12 +36,12 @@
     pre_launch_video_config,
     resolve_checkpoint_selector,
     resolve_published_checkpoint,
-    resolve_seed,
     run_playback,
     set_hydra_args,
     show_run_summary,
     startup_screen,
 )
+from . import cli_args_sb3 as cli_args
 
 logger = logging.getLogger(__name__)
 
@@ -105,9 +105,7 @@ def run(argv: list[str]) -> None:
         with launch_simulation(env_cfg, args_cli), contextlib.ExitStack() as cleanup:
             show_run_summary(screen, args_cli, env_cfg, library="sb3", action="play")
             apply_env_overrides(args_cli, env_cfg)
-            args_cli.seed = resolve_seed(args_cli.seed)
-            if args_cli.seed is not None:
-                agent_cfg["seed"] = args_cli.seed
+            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli)
             env_cfg.seed = agent_cfg["seed"]
 
             log_root_path = os.path.abspath(os.path.join("logs", "sb3", normalize_task_name(args_cli.task)))
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rl_games.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rl_games.py
index 88a6c95a94e5..3d503feb2a5b 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rl_games.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rl_games.py
@@ -41,13 +41,13 @@
     enable_cameras_for_video,
     pre_launch_video_config,
     resolve_checkpoint_selector,
-    resolve_seed,
     set_hydra_args,
     show_run_summary,
     startup_screen,
     wrap_sensor_capture,
     write_run_manifest,
 )
+from . import cli_args_rl_games as cli_args
 
 logger = logging.getLogger(__name__)
 
@@ -114,19 +114,10 @@ def run(argv: list[str]) -> None:
             show_run_summary(screen, args_cli, env_cfg, library="rl_games", action="train")
             apply_env_overrides(args_cli, env_cfg)
 
+            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli)
             params = agent_cfg["params"]
             config = params["config"]
-            args_cli.seed = resolve_seed(args_cli.seed)
-            if args_cli.seed is not None:
-                params["seed"] = args_cli.seed
-            if args_cli.max_iterations is not None:
-                config["max_epochs"] = args_cli.max_iterations
-            # the agent runs on the device launch_simulation resolved for the simulation
-            config["device"] = config["device_name"] = env_cfg.sim.device
             rank = int(os.getenv("RANK", "0")) if args_cli.distributed else None
-            if rank is not None:
-                params["seed"] += rank
-                config["multi_gpu"] = True
             env_cfg.seed = params["seed"]
 
             config_name = config["name"]
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rsl_rl.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rsl_rl.py
index 7b05d83ce1dc..d24a6f9d798f 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rsl_rl.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rsl_rl.py
@@ -124,8 +124,6 @@ def _run(args_cli: argparse.Namespace) -> None:
             apply_env_overrides(args_cli, env_cfg)
             if args_cli.max_iterations is not None:
                 agent_cfg.max_iterations = args_cli.max_iterations
-            # the agent runs on the device launch_simulation resolved for the simulation
-            agent_cfg.device = env_cfg.sim.device
             rank = int(os.getenv("RANK", "0")) if args_cli.distributed else None
             if rank is not None:
                 agent_cfg.seed += rank
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_sb3.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_sb3.py
index d7b085c99ca5..8c81957dce36 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_sb3.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_sb3.py
@@ -42,13 +42,13 @@
     enable_cameras_for_video,
     pre_launch_video_config,
     resolve_checkpoint_selector,
-    resolve_seed,
     set_hydra_args,
     show_run_summary,
     startup_screen,
     wrap_sensor_capture,
     write_run_manifest,
 )
+from . import cli_args_sb3 as cli_args
 
 logger = logging.getLogger(__name__)
 
@@ -115,14 +115,10 @@ def run(argv: list[str]) -> None:
         with launch_simulation(env_cfg, args_cli), contextlib.ExitStack() as cleanup:
             show_run_summary(screen, args_cli, env_cfg, library="sb3", action="train")
             apply_env_overrides(args_cli, env_cfg)
-            args_cli.seed = resolve_seed(args_cli.seed)
-            if args_cli.seed is not None:
-                agent_cfg["seed"] = args_cli.seed
+            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli)
             if args_cli.max_iterations is not None:
                 agent_cfg["n_timesteps"] = args_cli.max_iterations * agent_cfg["n_steps"] * env_cfg.scene.num_envs
             env_cfg.seed = agent_cfg["seed"]
-            # the agent runs on the device launch_simulation resolved for the simulation
-            agent_cfg["device"] = env_cfg.sim.device
 
             log_root_path = os.path.abspath(os.path.join("logs", "sb3", args_cli.task))
             logger.info(f"Logging experiment in directory: {log_root_path}")
diff --git a/source/isaaclab_rl/test/test_entrypoints_common.py b/source/isaaclab_rl/test/test_entrypoints_common.py
index 76396385ce02..95095be0da72 100644
--- a/source/isaaclab_rl/test/test_entrypoints_common.py
+++ b/source/isaaclab_rl/test/test_entrypoints_common.py
@@ -295,9 +295,9 @@ def test_run_summary_reports_the_launch_resolved_visualizers_and_device() -> Non
     """A run without visualizers shows headless, and the device is the one the launch resolved (e.g. per rank)."""
     from isaaclab.app import launch_simulation
 
-    sim_cfg = SimulationCfg(physics=NewtonCfg(), device="cuda:0")
+    sim_cfg = SimulationCfg(physics=NewtonCfg(), device="cpu")
     env_cfg = SimpleNamespace(sim=sim_cfg, scene=SimpleNamespace(num_envs=4))
-    args_cli = argparse.Namespace(task="Isaac-Test", device="cuda:0", num_envs=None, visualizer=None)
+    args_cli = argparse.Namespace(task="Isaac-Test", device="cpu", num_envs=None, visualizer=None)
     screen = _RecordingScreen()
 
     with launch_simulation(env_cfg, args_cli):

From 062dd2adc443ad16832094830a9180a483c11016 Mon Sep 17 00:00:00 2001
From: Antoine RICHARD <antoiner@nvidia.com>
Date: Mon, 5 Oct 2026 09:56:16 +0200
Subject: [PATCH 6/9] Pass the agent device and seed to the shared RL config
 setup explicitly

---
 .../rl_games/benchmark_play_rl_games.py       |  2 +-
 .../rl_games/benchmark_train_rl_games.py      | 11 +++-
 .../backends/rsl_rl/benchmark_play_rsl_rl.py  |  2 +-
 .../backends/rsl_rl/benchmark_train_rsl_rl.py | 11 +++-
 .../backends/sb3/benchmark_play_sb3.py        |  4 +-
 .../backends/sb3/benchmark_train_sb3.py       |  2 +-
 ...fix-rsl-rl-runner-cpu-fallback.changed.rst | 11 +---
 .../entrypoints/backends/cli_args_rl_games.py | 28 +++-----
 .../entrypoints/backends/cli_args_rsl_rl.py   | 64 +++++++++++++------
 .../entrypoints/backends/cli_args_sb3.py      | 18 ++----
 .../entrypoints/backends/export_rl_games.py   |  2 +-
 .../entrypoints/backends/export_rsl_rl.py     |  8 ++-
 .../entrypoints/backends/play_rl_games.py     |  2 +-
 .../entrypoints/backends/play_rsl_rl.py       | 11 +++-
 .../entrypoints/backends/play_sb3.py          |  4 +-
 .../entrypoints/backends/train_rl_games.py    |  7 +-
 .../entrypoints/backends/train_rsl_rl.py      | 11 +++-
 .../entrypoints/backends/train_sb3.py         |  2 +-
 18 files changed, 124 insertions(+), 76 deletions(-)

diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_play_rl_games.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_play_rl_games.py
index e928c9ec9fde..b251cfa5e3a4 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_play_rl_games.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_play_rl_games.py
@@ -140,7 +140,7 @@ def run(argv: list[str]) -> BenchmarkResult:
 
             if args_cli.num_envs is not None:
                 env_cfg.scene.num_envs = args_cli.num_envs
-            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli)
+            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli.device, seed=args_cli.seed)
             env_cfg.seed = agent_cfg["params"]["seed"]
 
             config_name = agent_cfg["params"]["config"]["name"]
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_train_rl_games.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_train_rl_games.py
index 1582cfd45887..03b0dfadf2df 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_train_rl_games.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_train_rl_games.py
@@ -194,7 +194,16 @@ def run(argv: list[str]) -> BenchmarkResult | None:
 
             apply_env_overrides(args_cli, env_cfg)
 
-            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli)
+            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli.device, seed=args_cli.seed)
+
+            if args_cli.max_iterations is not None:
+                agent_cfg["params"]["config"]["max_epochs"] = args_cli.max_iterations
+
+            if distributed.enabled:
+                # Mirror the regular training entrypoint: offsetting the seed by the rank decorrelates
+                # exploration across ranks.
+                agent_cfg["params"]["seed"] += distributed.rank
+                agent_cfg["params"]["config"]["multi_gpu"] = True
             env_cfg.seed = agent_cfg["params"]["seed"]
             horizon_length = agent_cfg["params"]["config"].get("horizon_length", 16)
             reported_num_envs, _ = distributed.global_work(env_cfg.scene.num_envs, horizon_length)
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_play_rsl_rl.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_play_rsl_rl.py
index c5a5a2b53a25..61e05f3fdaa9 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_play_rsl_rl.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_play_rsl_rl.py
@@ -138,7 +138,7 @@ def run(argv: list[str]) -> BenchmarkResult:
 
             if args.num_envs is not None:
                 env_cfg.scene.num_envs = args.num_envs
-            agent_cfg = cli_args.update_rsl_rl_cfg(agent_cfg, args)
+            agent_cfg = cli_args.update_rsl_rl_cfg(agent_cfg, args.device, seed=args.seed)
             env_cfg.seed = agent_cfg.seed
 
             log_root_path = os.path.abspath(os.path.join("logs", "rsl_rl", agent_cfg.experiment_name))
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_train_rsl_rl.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_train_rsl_rl.py
index dfd5f2af1e61..b412c74f456c 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_train_rsl_rl.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_train_rsl_rl.py
@@ -185,7 +185,16 @@ def run(argv: list[str]) -> BenchmarkResult | None:
             app_t1 = time.perf_counter_ns()
 
             apply_env_overrides(args_cli, env_cfg)
-            agent_cfg = cli_args.update_rsl_rl_cfg(agent_cfg, args_cli)
+            agent_cfg = cli_args.update_rsl_rl_cfg(
+                agent_cfg,
+                args_cli.device,
+                seed=args_cli.seed,
+                checkpoint=args_cli.checkpoint,
+                experiment_name=args_cli.experiment_name,
+                run_name=args_cli.run_name,
+                logger=args_cli.logger,
+                log_project_name=args_cli.log_project_name,
+            )
             agent_cfg.max_iterations = (
                 args_cli.max_iterations if args_cli.max_iterations is not None else agent_cfg.max_iterations
             )
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_play_sb3.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_play_sb3.py
index 8d4671cc7415..b8dc85631d0f 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_play_sb3.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_play_sb3.py
@@ -143,7 +143,7 @@ def run(argv: list[str]) -> BenchmarkResult:
 
             if args_cli.num_envs is not None:
                 env_cfg.scene.num_envs = args_cli.num_envs
-            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli)
+            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli.device, seed=args_cli.seed)
             env_cfg.seed = agent_cfg["seed"]
 
             log_root_path = os.path.abspath(os.path.join("logs", "sb3", args_cli.task))
@@ -213,7 +213,7 @@ def run(argv: list[str]) -> BenchmarkResult:
                 )
 
             # Load the trained policy.
-            agent = PPO.load(resume_path, env, device=env_cfg.sim.device, print_system_info=True)
+            agent = PPO.load(resume_path, env, device=agent_cfg["device"], print_system_info=True)
 
             def policy(obs):
                 """Map an observation batch to a deterministic action batch via the sb3 agent.
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_train_sb3.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_train_sb3.py
index 7b23c0cfb17f..090cdf9a856c 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_train_sb3.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_train_sb3.py
@@ -241,7 +241,7 @@ def run(argv: list[str]) -> BenchmarkResult:
 
             apply_env_overrides(args_cli, env_cfg)
 
-            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli)
+            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli.device, seed=args_cli.seed)
             env_cfg.seed = agent_cfg["seed"]
 
             # Convert the iteration override to SB3 total timesteps.
diff --git a/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst b/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst
index dec9ccf74c4c..fabe2c924fa1 100644
--- a/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst
+++ b/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst
@@ -1,8 +1,3 @@
-* **Breaking:** Changed the RSL-RL, RL-Games, and Stable-Baselines3 train, play, and export workflows to run the
-  agent on the device :func:`~isaaclab.app.launch_simulation` resolved for the simulation, as skrl and TorchRL
-  already did. Outside distributed runs they used the agent configuration's device, ``cuda:0`` by default, so
-  ``--device cpu`` trained on ``cuda:0`` and ``--device cuda:1`` trained on GPU 0. Pass ``--device`` to choose
-  where both the simulation and the agent run; the agent configuration's ``device`` is now overwritten.
-* Changed each library's agent configuration overrides to live in one shared function that every train, play,
-  export, and benchmark workflow calls: ``update_rsl_rl_cfg``, and the new ``update_rl_games_cfg`` and
-  ``update_sb3_cfg``.
+* **Breaking:** Changed the RSL-RL, RL-Games, and Stable-Baselines3 workflows to run the agent on the simulation's
+  device, as skrl and TorchRL do. They ignored ``--device`` for the agent outside distributed runs and used the
+  agent configuration's ``cuda:0``; pass ``--device`` to choose where both run.
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rl_games.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rl_games.py
index a85d04135d32..373368be2746 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rl_games.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rl_games.py
@@ -7,35 +7,23 @@
 
 from __future__ import annotations
 
-import argparse
-import os
-
 from ..common import resolve_seed
 
 
-def update_rl_games_cfg(agent_cfg: dict, args_cli: argparse.Namespace) -> dict:
+def update_rl_games_cfg(agent_cfg: dict, device: str, *, seed: int | None = None) -> dict:
     """Override an RL-Games agent configuration with the command-line arguments.
 
-    Call it inside :func:`~isaaclab.app.launch_simulation`, which resolves ``args_cli.device``: the agent runs on
-    the simulation's device.
-
     Args:
         agent_cfg: The configuration for RL-Games agent.
-        args_cli: The command line arguments.
+        device: Device the agent runs on, the simulation's device once :func:`~isaaclab.app.launch_simulation`
+            has resolved it.
+        seed: Seed override; ``-1`` draws a random seed.
 
     Returns:
         The updated RL-Games agent configuration.
     """
-    params = agent_cfg["params"]
-    config = params["config"]
-    config["device"] = config["device_name"] = args_cli.device
-    if getattr(args_cli, "seed", None) is not None:
-        args_cli.seed = resolve_seed(args_cli.seed)
-        params["seed"] = args_cli.seed
-    if getattr(args_cli, "max_iterations", None) is not None:
-        config["max_epochs"] = args_cli.max_iterations
-    if getattr(args_cli, "distributed", False):
-        # offsetting the seed by the rank decorrelates exploration across ranks
-        params["seed"] += int(os.getenv("RANK", "0"))
-        config["multi_gpu"] = True
+    config = agent_cfg["params"]["config"]
+    config["device"] = config["device_name"] = device
+    if seed is not None:
+        agent_cfg["params"]["seed"] = resolve_seed(seed)
     return agent_cfg
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rsl_rl.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rsl_rl.py
index 484858cff748..9c461453c6df 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rsl_rl.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rsl_rl.py
@@ -81,35 +81,57 @@ def parse_rsl_rl_cfg(task_name: str, args_cli: argparse.Namespace) -> RslRlBaseR
     from isaaclab_tasks.utils.parse_cfg import load_cfg_from_registry
 
     agent_cfg: RslRlBaseRunnerCfg = load_cfg_from_registry(task_name, "rsl_rl_cfg_entry_point")
-    return update_rsl_rl_cfg(agent_cfg, args_cli)
+    return update_rsl_rl_cfg(
+        agent_cfg,
+        args_cli.device,
+        seed=args_cli.seed,
+        checkpoint=args_cli.checkpoint,
+        experiment_name=args_cli.experiment_name,
+        run_name=args_cli.run_name,
+        logger=args_cli.logger,
+        log_project_name=args_cli.log_project_name,
+    )
 
 
-def update_rsl_rl_cfg(agent_cfg: RslRlBaseRunnerCfg, args_cli: argparse.Namespace) -> RslRlBaseRunnerCfg:
+def update_rsl_rl_cfg(
+    agent_cfg: RslRlBaseRunnerCfg,
+    device: str,
+    *,
+    seed: int | None = None,
+    checkpoint: str | None = None,
+    experiment_name: str | None = None,
+    run_name: str | None = None,
+    logger: str | None = None,
+    log_project_name: str | None = None,
+) -> RslRlBaseRunnerCfg:
     """Override an RSL-RL agent configuration with the command-line arguments.
 
-    Call it inside :func:`~isaaclab.app.launch_simulation`, which resolves ``args_cli.device``: the agent runs on
-    the simulation's device.
-
     Args:
         agent_cfg: The configuration for RSL-RL agent.
-        args_cli: The command line arguments.
+        device: Device the agent runs on, the simulation's device once :func:`~isaaclab.app.launch_simulation`
+            has resolved it.
+        seed: Seed override; ``-1`` draws a random seed.
+        checkpoint: Checkpoint file to load.
+        experiment_name: Experiment folder name.
+        run_name: Run name suffix.
+        logger: Logger to use.
+        log_project_name: Project name for the ``wandb`` and ``neptune`` loggers.
 
     Returns:
         The updated RSL-RL agent configuration.
     """
-    agent_cfg.device = args_cli.device
-    if getattr(args_cli, "seed", None) is not None:
-        args_cli.seed = resolve_seed(args_cli.seed)
-        agent_cfg.seed = args_cli.seed
-    if getattr(args_cli, "checkpoint", None) is not None:
-        agent_cfg.load_checkpoint = args_cli.checkpoint
-    if getattr(args_cli, "experiment_name", None) is not None:
-        agent_cfg.experiment_name = args_cli.experiment_name
-    if getattr(args_cli, "run_name", None) is not None:
-        agent_cfg.run_name = args_cli.run_name
-    if getattr(args_cli, "logger", None) is not None:
-        agent_cfg.logger = args_cli.logger
-    if agent_cfg.logger in {"wandb", "neptune"} and getattr(args_cli, "log_project_name", None):
-        agent_cfg.wandb_project = args_cli.log_project_name
-        agent_cfg.neptune_project = args_cli.log_project_name
+    agent_cfg.device = device
+    if seed is not None:
+        agent_cfg.seed = resolve_seed(seed)
+    if checkpoint is not None:
+        agent_cfg.load_checkpoint = checkpoint
+    if experiment_name is not None:
+        agent_cfg.experiment_name = experiment_name
+    if run_name is not None:
+        agent_cfg.run_name = run_name
+    if logger is not None:
+        agent_cfg.logger = logger
+    if agent_cfg.logger in {"wandb", "neptune"} and log_project_name:
+        agent_cfg.wandb_project = log_project_name
+        agent_cfg.neptune_project = log_project_name
     return agent_cfg
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_sb3.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_sb3.py
index 71f3d2cb8cf0..658e28dec74a 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_sb3.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_sb3.py
@@ -7,26 +7,22 @@
 
 from __future__ import annotations
 
-import argparse
-
 from ..common import resolve_seed
 
 
-def update_sb3_cfg(agent_cfg: dict, args_cli: argparse.Namespace) -> dict:
+def update_sb3_cfg(agent_cfg: dict, device: str, *, seed: int | None = None) -> dict:
     """Override a Stable-Baselines3 agent configuration with the command-line arguments.
 
-    Call it inside :func:`~isaaclab.app.launch_simulation`, which resolves ``args_cli.device``: the agent runs on
-    the simulation's device. Loading a checkpoint still needs that device passed to ``PPO.load``.
-
     Args:
         agent_cfg: The configuration for Stable-Baselines3 agent.
-        args_cli: The command line arguments.
+        device: Device the agent runs on, the simulation's device once :func:`~isaaclab.app.launch_simulation`
+            has resolved it. Pass ``agent_cfg["device"]`` to ``PPO.load`` as well.
+        seed: Seed override; ``-1`` draws a random seed.
 
     Returns:
         The updated Stable-Baselines3 agent configuration.
     """
-    agent_cfg["device"] = args_cli.device
-    if getattr(args_cli, "seed", None) is not None:
-        args_cli.seed = resolve_seed(args_cli.seed)
-        agent_cfg["seed"] = args_cli.seed
+    agent_cfg["device"] = device
+    if seed is not None:
+        agent_cfg["seed"] = resolve_seed(seed)
     return agent_cfg
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rl_games.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rl_games.py
index 7601ceb1be0f..1d9fe4faa208 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rl_games.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rl_games.py
@@ -106,7 +106,7 @@ def export_rl_games_agent(args_cli: argparse.Namespace, env_cfg: Any, agent_cfg:
     # initialized the selected backend
     from leapp import annotate
 
-    agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli)
+    agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli.device)
     params = agent_cfg["params"]
     env_cfg.scene.num_envs = 1
     env_cfg.seed = params["seed"]
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rsl_rl.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rsl_rl.py
index e1fcfd329bb7..74be73564f07 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rsl_rl.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rsl_rl.py
@@ -132,7 +132,13 @@ def export_rsl_rl_agent(args_cli: argparse.Namespace, env_cfg: Any, agent_cfg: R
     # the LEAPP runtime loads simulation modules, so import it only after the launch
     from leapp import annotate
 
-    agent_cfg = cli_args.update_rsl_rl_cfg(agent_cfg, args_cli)
+    agent_cfg = cli_args.update_rsl_rl_cfg(
+        agent_cfg,
+        args_cli.device,
+        seed=args_cli.seed,
+        checkpoint=args_cli.checkpoint,
+        experiment_name=args_cli.experiment_name,
+    )
 
     env_cfg.scene.num_envs = 1
     # certain randomizations occur in the environment initialization so we set the seed here
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rl_games.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rl_games.py
index b4376e0356e1..4cc727222518 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rl_games.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rl_games.py
@@ -112,7 +112,7 @@ def run(argv: list[str]) -> None:
         with launch_simulation(env_cfg, args_cli), contextlib.ExitStack() as cleanup:
             show_run_summary(screen, args_cli, env_cfg, library="rl_games", action="play")
             apply_env_overrides(args_cli, env_cfg)
-            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli)
+            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli.device, seed=args_cli.seed)
             params = agent_cfg["params"]
             env_cfg.seed = params["seed"]
 
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rsl_rl.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rsl_rl.py
index 46c32aceef74..ac8911616d9e 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rsl_rl.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rsl_rl.py
@@ -107,7 +107,16 @@ def run(argv: list[str]) -> None:
         screen.stage("Launching simulation")
         with launch_simulation(env_cfg, args_cli), contextlib.ExitStack() as cleanup:
             show_run_summary(screen, args_cli, env_cfg, library="rsl_rl", action="play")
-            agent_cfg = cli_args.update_rsl_rl_cfg(agent_cfg, args_cli)
+            agent_cfg = cli_args.update_rsl_rl_cfg(
+                agent_cfg,
+                args_cli.device,
+                seed=args_cli.seed,
+                checkpoint=args_cli.checkpoint,
+                experiment_name=args_cli.experiment_name,
+                run_name=args_cli.run_name,
+                logger=args_cli.logger,
+                log_project_name=args_cli.log_project_name,
+            )
             apply_env_overrides(args_cli, env_cfg)
             # certain randomizations occur in the environment initialization so we set the seed here
             env_cfg.seed = agent_cfg.seed
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_sb3.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_sb3.py
index 3371ffd8e936..d6669241cf27 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_sb3.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_sb3.py
@@ -105,7 +105,7 @@ def run(argv: list[str]) -> None:
         with launch_simulation(env_cfg, args_cli), contextlib.ExitStack() as cleanup:
             show_run_summary(screen, args_cli, env_cfg, library="sb3", action="play")
             apply_env_overrides(args_cli, env_cfg)
-            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli)
+            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli.device, seed=args_cli.seed)
             env_cfg.seed = agent_cfg["seed"]
 
             log_root_path = os.path.abspath(os.path.join("logs", "sb3", normalize_task_name(args_cli.task)))
@@ -145,7 +145,7 @@ def run(argv: list[str]) -> None:
                 )
 
             print(f"Loading checkpoint from: {checkpoint_path}")
-            agent = PPO.load(checkpoint_path, env, device=env_cfg.sim.device, print_system_info=True)
+            agent = PPO.load(checkpoint_path, env, device=agent_cfg["device"], print_system_info=True)
             # configure_seed must run after PPO.load so torch determinism does not disturb SB3's initialization
             if args_cli.deterministic:
                 configure_seed(env_cfg.seed, torch_deterministic=True)
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rl_games.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rl_games.py
index 3d503feb2a5b..325f39f4451e 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rl_games.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rl_games.py
@@ -114,10 +114,15 @@ def run(argv: list[str]) -> None:
             show_run_summary(screen, args_cli, env_cfg, library="rl_games", action="train")
             apply_env_overrides(args_cli, env_cfg)
 
-            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli)
+            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli.device, seed=args_cli.seed)
             params = agent_cfg["params"]
             config = params["config"]
+            if args_cli.max_iterations is not None:
+                config["max_epochs"] = args_cli.max_iterations
             rank = int(os.getenv("RANK", "0")) if args_cli.distributed else None
+            if rank is not None:
+                params["seed"] += rank
+                config["multi_gpu"] = True
             env_cfg.seed = params["seed"]
 
             config_name = config["name"]
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rsl_rl.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rsl_rl.py
index d24a6f9d798f..668128fabce6 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rsl_rl.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rsl_rl.py
@@ -120,7 +120,16 @@ def _run(args_cli: argparse.Namespace) -> None:
         screen.stage("Launching simulation")
         with launch_simulation(env_cfg, args_cli), contextlib.ExitStack() as cleanup:
             show_run_summary(screen, args_cli, env_cfg, library="rsl_rl", action="train")
-            agent_cfg = cli_args.update_rsl_rl_cfg(agent_cfg, args_cli)
+            agent_cfg = cli_args.update_rsl_rl_cfg(
+                agent_cfg,
+                args_cli.device,
+                seed=args_cli.seed,
+                checkpoint=args_cli.checkpoint,
+                experiment_name=args_cli.experiment_name,
+                run_name=args_cli.run_name,
+                logger=args_cli.logger,
+                log_project_name=args_cli.log_project_name,
+            )
             apply_env_overrides(args_cli, env_cfg)
             if args_cli.max_iterations is not None:
                 agent_cfg.max_iterations = args_cli.max_iterations
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_sb3.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_sb3.py
index 8c81957dce36..b79dd093b6e3 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_sb3.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_sb3.py
@@ -115,7 +115,7 @@ def run(argv: list[str]) -> None:
         with launch_simulation(env_cfg, args_cli), contextlib.ExitStack() as cleanup:
             show_run_summary(screen, args_cli, env_cfg, library="sb3", action="train")
             apply_env_overrides(args_cli, env_cfg)
-            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli)
+            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli.device, seed=args_cli.seed)
             if args_cli.max_iterations is not None:
                 agent_cfg["n_timesteps"] = args_cli.max_iterations * agent_cfg["n_steps"] * env_cfg.scene.num_envs
             env_cfg.seed = agent_cfg["seed"]

From 9d208ef7b62dae5c9f1b0ff993b74bb5034c206e Mon Sep 17 00:00:00 2001
From: Antoine RICHARD <antoiner@nvidia.com>
Date: Mon, 5 Oct 2026 10:55:51 +0200
Subject: [PATCH 7/9] Restore update_rsl_rl_cfg's args_cli signature and set
 the device there

---
 .../backends/rsl_rl/benchmark_play_rsl_rl.py  |  5 +-
 .../backends/rsl_rl/benchmark_train_rsl_rl.py | 11 +---
 .../entrypoints/backends/cli_args_rsl_rl.py   | 61 ++++++-------------
 .../entrypoints/backends/export_rsl_rl.py     | 17 +++---
 .../entrypoints/backends/play_rsl_rl.py       | 11 +---
 .../entrypoints/backends/train_rsl_rl.py      | 11 +---
 6 files changed, 32 insertions(+), 84 deletions(-)

diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_play_rsl_rl.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_play_rsl_rl.py
index 61e05f3fdaa9..418154991ab5 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_play_rsl_rl.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_play_rsl_rl.py
@@ -111,7 +111,6 @@ def run(argv: list[str]) -> BenchmarkResult:
     from isaaclab.benchmark import BaseIsaacLabBenchmark, BenchmarkMonitor, BenchmarkResult, builders, capture, stepping
     from isaaclab.benchmark.schema import StartupTime
 
-    from isaaclab_rl.entrypoints.backends import cli_args_rsl_rl as cli_args
     from isaaclab_rl.rsl_rl import RslRlVecEnvWrapper
 
     # Importing the task packages registers their gym environments so the
@@ -138,7 +137,9 @@ def run(argv: list[str]) -> BenchmarkResult:
 
             if args.num_envs is not None:
                 env_cfg.scene.num_envs = args.num_envs
-            agent_cfg = cli_args.update_rsl_rl_cfg(agent_cfg, args.device, seed=args.seed)
+            if args.seed is not None:
+                agent_cfg.seed = args.seed
+            agent_cfg.device = args.device
             env_cfg.seed = agent_cfg.seed
 
             log_root_path = os.path.abspath(os.path.join("logs", "rsl_rl", agent_cfg.experiment_name))
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_train_rsl_rl.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_train_rsl_rl.py
index b412c74f456c..dfd5f2af1e61 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_train_rsl_rl.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_train_rsl_rl.py
@@ -185,16 +185,7 @@ def run(argv: list[str]) -> BenchmarkResult | None:
             app_t1 = time.perf_counter_ns()
 
             apply_env_overrides(args_cli, env_cfg)
-            agent_cfg = cli_args.update_rsl_rl_cfg(
-                agent_cfg,
-                args_cli.device,
-                seed=args_cli.seed,
-                checkpoint=args_cli.checkpoint,
-                experiment_name=args_cli.experiment_name,
-                run_name=args_cli.run_name,
-                logger=args_cli.logger,
-                log_project_name=args_cli.log_project_name,
-            )
+            agent_cfg = cli_args.update_rsl_rl_cfg(agent_cfg, args_cli)
             agent_cfg.max_iterations = (
                 args_cli.max_iterations if args_cli.max_iterations is not None else agent_cfg.max_iterations
             )
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rsl_rl.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rsl_rl.py
index 9c461453c6df..df9996093480 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rsl_rl.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rsl_rl.py
@@ -81,57 +81,32 @@ def parse_rsl_rl_cfg(task_name: str, args_cli: argparse.Namespace) -> RslRlBaseR
     from isaaclab_tasks.utils.parse_cfg import load_cfg_from_registry
 
     agent_cfg: RslRlBaseRunnerCfg = load_cfg_from_registry(task_name, "rsl_rl_cfg_entry_point")
-    return update_rsl_rl_cfg(
-        agent_cfg,
-        args_cli.device,
-        seed=args_cli.seed,
-        checkpoint=args_cli.checkpoint,
-        experiment_name=args_cli.experiment_name,
-        run_name=args_cli.run_name,
-        logger=args_cli.logger,
-        log_project_name=args_cli.log_project_name,
-    )
+    return update_rsl_rl_cfg(agent_cfg, args_cli)
 
 
-def update_rsl_rl_cfg(
-    agent_cfg: RslRlBaseRunnerCfg,
-    device: str,
-    *,
-    seed: int | None = None,
-    checkpoint: str | None = None,
-    experiment_name: str | None = None,
-    run_name: str | None = None,
-    logger: str | None = None,
-    log_project_name: str | None = None,
-) -> RslRlBaseRunnerCfg:
+def update_rsl_rl_cfg(agent_cfg: RslRlBaseRunnerCfg, args_cli: argparse.Namespace) -> RslRlBaseRunnerCfg:
     """Override an RSL-RL agent configuration with the command-line arguments.
 
     Args:
         agent_cfg: The configuration for RSL-RL agent.
-        device: Device the agent runs on, the simulation's device once :func:`~isaaclab.app.launch_simulation`
-            has resolved it.
-        seed: Seed override; ``-1`` draws a random seed.
-        checkpoint: Checkpoint file to load.
-        experiment_name: Experiment folder name.
-        run_name: Run name suffix.
-        logger: Logger to use.
-        log_project_name: Project name for the ``wandb`` and ``neptune`` loggers.
+        args_cli: The command line arguments.
 
     Returns:
         The updated RSL-RL agent configuration.
     """
-    agent_cfg.device = device
-    if seed is not None:
-        agent_cfg.seed = resolve_seed(seed)
-    if checkpoint is not None:
-        agent_cfg.load_checkpoint = checkpoint
-    if experiment_name is not None:
-        agent_cfg.experiment_name = experiment_name
-    if run_name is not None:
-        agent_cfg.run_name = run_name
-    if logger is not None:
-        agent_cfg.logger = logger
-    if agent_cfg.logger in {"wandb", "neptune"} and log_project_name:
-        agent_cfg.wandb_project = log_project_name
-        agent_cfg.neptune_project = log_project_name
+    agent_cfg.device = args_cli.device
+    if getattr(args_cli, "seed", None) is not None:
+        args_cli.seed = resolve_seed(args_cli.seed)
+        agent_cfg.seed = args_cli.seed
+    if args_cli.checkpoint is not None:
+        agent_cfg.load_checkpoint = args_cli.checkpoint
+    if args_cli.experiment_name is not None:
+        agent_cfg.experiment_name = args_cli.experiment_name
+    if args_cli.run_name is not None:
+        agent_cfg.run_name = args_cli.run_name
+    if args_cli.logger is not None:
+        agent_cfg.logger = args_cli.logger
+    if agent_cfg.logger in {"wandb", "neptune"} and args_cli.log_project_name:
+        agent_cfg.wandb_project = args_cli.log_project_name
+        agent_cfg.neptune_project = args_cli.log_project_name
     return agent_cfg
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rsl_rl.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rsl_rl.py
index 74be73564f07..baf46a74c11a 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rsl_rl.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rsl_rl.py
@@ -30,8 +30,7 @@
 import isaaclab_tasks  # noqa: F401
 
 from ...rsl_rl import RslRlBaseRunnerCfg, RslRlVecEnvWrapper
-from ..common import resolve_published_checkpoint
-from . import cli_args_rsl_rl as cli_args
+from ..common import resolve_published_checkpoint, resolve_seed
 from .export_common import (
     add_common_export_args,
     finalize_export_args,
@@ -132,13 +131,13 @@ def export_rsl_rl_agent(args_cli: argparse.Namespace, env_cfg: Any, agent_cfg: R
     # the LEAPP runtime loads simulation modules, so import it only after the launch
     from leapp import annotate
 
-    agent_cfg = cli_args.update_rsl_rl_cfg(
-        agent_cfg,
-        args_cli.device,
-        seed=args_cli.seed,
-        checkpoint=args_cli.checkpoint,
-        experiment_name=args_cli.experiment_name,
-    )
+    if args_cli.seed is not None:
+        agent_cfg.seed = resolve_seed(args_cli.seed)
+    if args_cli.checkpoint is not None:
+        agent_cfg.load_checkpoint = args_cli.checkpoint
+    if args_cli.experiment_name is not None:
+        agent_cfg.experiment_name = args_cli.experiment_name
+    agent_cfg.device = args_cli.device
 
     env_cfg.scene.num_envs = 1
     # certain randomizations occur in the environment initialization so we set the seed here
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rsl_rl.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rsl_rl.py
index ac8911616d9e..46c32aceef74 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rsl_rl.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rsl_rl.py
@@ -107,16 +107,7 @@ def run(argv: list[str]) -> None:
         screen.stage("Launching simulation")
         with launch_simulation(env_cfg, args_cli), contextlib.ExitStack() as cleanup:
             show_run_summary(screen, args_cli, env_cfg, library="rsl_rl", action="play")
-            agent_cfg = cli_args.update_rsl_rl_cfg(
-                agent_cfg,
-                args_cli.device,
-                seed=args_cli.seed,
-                checkpoint=args_cli.checkpoint,
-                experiment_name=args_cli.experiment_name,
-                run_name=args_cli.run_name,
-                logger=args_cli.logger,
-                log_project_name=args_cli.log_project_name,
-            )
+            agent_cfg = cli_args.update_rsl_rl_cfg(agent_cfg, args_cli)
             apply_env_overrides(args_cli, env_cfg)
             # certain randomizations occur in the environment initialization so we set the seed here
             env_cfg.seed = agent_cfg.seed
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rsl_rl.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rsl_rl.py
index 668128fabce6..d24a6f9d798f 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rsl_rl.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rsl_rl.py
@@ -120,16 +120,7 @@ def _run(args_cli: argparse.Namespace) -> None:
         screen.stage("Launching simulation")
         with launch_simulation(env_cfg, args_cli), contextlib.ExitStack() as cleanup:
             show_run_summary(screen, args_cli, env_cfg, library="rsl_rl", action="train")
-            agent_cfg = cli_args.update_rsl_rl_cfg(
-                agent_cfg,
-                args_cli.device,
-                seed=args_cli.seed,
-                checkpoint=args_cli.checkpoint,
-                experiment_name=args_cli.experiment_name,
-                run_name=args_cli.run_name,
-                logger=args_cli.logger,
-                log_project_name=args_cli.log_project_name,
-            )
+            agent_cfg = cli_args.update_rsl_rl_cfg(agent_cfg, args_cli)
             apply_env_overrides(args_cli, env_cfg)
             if args_cli.max_iterations is not None:
                 agent_cfg.max_iterations = args_cli.max_iterations

From 202545fa937e9a6251c496b3e8813278cd96ad84 Mon Sep 17 00:00:00 2001
From: Antoine RICHARD <antoiner@nvidia.com>
Date: Mon, 5 Oct 2026 11:01:37 +0200
Subject: [PATCH 8/9] Pass args_cli to update_sb3_cfg like update_rsl_rl_cfg

---
 .../backends/sb3/benchmark_play_sb3.py            |  2 +-
 .../backends/sb3/benchmark_train_sb3.py           |  2 +-
 .../entrypoints/backends/cli_args_sb3.py          | 15 ++++++++-------
 .../isaaclab_rl/entrypoints/backends/play_sb3.py  |  2 +-
 .../isaaclab_rl/entrypoints/backends/train_sb3.py |  2 +-
 5 files changed, 12 insertions(+), 11 deletions(-)

diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_play_sb3.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_play_sb3.py
index b8dc85631d0f..cdd296a7863e 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_play_sb3.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_play_sb3.py
@@ -143,7 +143,7 @@ def run(argv: list[str]) -> BenchmarkResult:
 
             if args_cli.num_envs is not None:
                 env_cfg.scene.num_envs = args_cli.num_envs
-            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli.device, seed=args_cli.seed)
+            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli)
             env_cfg.seed = agent_cfg["seed"]
 
             log_root_path = os.path.abspath(os.path.join("logs", "sb3", args_cli.task))
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_train_sb3.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_train_sb3.py
index 090cdf9a856c..7b23c0cfb17f 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_train_sb3.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/sb3/benchmark_train_sb3.py
@@ -241,7 +241,7 @@ def run(argv: list[str]) -> BenchmarkResult:
 
             apply_env_overrides(args_cli, env_cfg)
 
-            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli.device, seed=args_cli.seed)
+            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli)
             env_cfg.seed = agent_cfg["seed"]
 
             # Convert the iteration override to SB3 total timesteps.
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_sb3.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_sb3.py
index 658e28dec74a..25e44c3df8a8 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_sb3.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_sb3.py
@@ -7,22 +7,23 @@
 
 from __future__ import annotations
 
+import argparse
+
 from ..common import resolve_seed
 
 
-def update_sb3_cfg(agent_cfg: dict, device: str, *, seed: int | None = None) -> dict:
+def update_sb3_cfg(agent_cfg: dict, args_cli: argparse.Namespace) -> dict:
     """Override a Stable-Baselines3 agent configuration with the command-line arguments.
 
     Args:
         agent_cfg: The configuration for Stable-Baselines3 agent.
-        device: Device the agent runs on, the simulation's device once :func:`~isaaclab.app.launch_simulation`
-            has resolved it. Pass ``agent_cfg["device"]`` to ``PPO.load`` as well.
-        seed: Seed override; ``-1`` draws a random seed.
+        args_cli: The command line arguments.
 
     Returns:
         The updated Stable-Baselines3 agent configuration.
     """
-    agent_cfg["device"] = device
-    if seed is not None:
-        agent_cfg["seed"] = resolve_seed(seed)
+    agent_cfg["device"] = args_cli.device
+    if args_cli.seed is not None:
+        args_cli.seed = resolve_seed(args_cli.seed)
+        agent_cfg["seed"] = args_cli.seed
     return agent_cfg
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_sb3.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_sb3.py
index d6669241cf27..150abc2fa724 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_sb3.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_sb3.py
@@ -105,7 +105,7 @@ def run(argv: list[str]) -> None:
         with launch_simulation(env_cfg, args_cli), contextlib.ExitStack() as cleanup:
             show_run_summary(screen, args_cli, env_cfg, library="sb3", action="play")
             apply_env_overrides(args_cli, env_cfg)
-            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli.device, seed=args_cli.seed)
+            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli)
             env_cfg.seed = agent_cfg["seed"]
 
             log_root_path = os.path.abspath(os.path.join("logs", "sb3", normalize_task_name(args_cli.task)))
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_sb3.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_sb3.py
index b79dd093b6e3..8c81957dce36 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_sb3.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_sb3.py
@@ -115,7 +115,7 @@ def run(argv: list[str]) -> None:
         with launch_simulation(env_cfg, args_cli), contextlib.ExitStack() as cleanup:
             show_run_summary(screen, args_cli, env_cfg, library="sb3", action="train")
             apply_env_overrides(args_cli, env_cfg)
-            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli.device, seed=args_cli.seed)
+            agent_cfg = cli_args.update_sb3_cfg(agent_cfg, args_cli)
             if args_cli.max_iterations is not None:
                 agent_cfg["n_timesteps"] = args_cli.max_iterations * agent_cfg["n_steps"] * env_cfg.scene.num_envs
             env_cfg.seed = agent_cfg["seed"]

From 9055fc6c522e45d7e9f0ffdfffb5a197be87fec2 Mon Sep 17 00:00:00 2001
From: Antoine RICHARD <antoiner@nvidia.com>
Date: Mon, 5 Oct 2026 11:09:58 +0200
Subject: [PATCH 9/9] Add --seed to RL-Games export and pass args_cli to
 update_rl_games_cfg

---
 .../backends/rl_games/benchmark_play_rl_games.py  |  2 +-
 .../backends/rl_games/benchmark_train_rl_games.py |  2 +-
 ...oiner-fix-rsl-rl-runner-cpu-fallback.added.rst |  1 +
 .../entrypoints/backends/cli_args_rl_games.py     | 15 ++++++++-------
 .../entrypoints/backends/export_rl_games.py       |  3 ++-
 .../entrypoints/backends/play_rl_games.py         |  2 +-
 .../entrypoints/backends/train_rl_games.py        |  2 +-
 7 files changed, 15 insertions(+), 12 deletions(-)
 create mode 100644 source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.added.rst

diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_play_rl_games.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_play_rl_games.py
index b251cfa5e3a4..e928c9ec9fde 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_play_rl_games.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_play_rl_games.py
@@ -140,7 +140,7 @@ def run(argv: list[str]) -> BenchmarkResult:
 
             if args_cli.num_envs is not None:
                 env_cfg.scene.num_envs = args_cli.num_envs
-            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli.device, seed=args_cli.seed)
+            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli)
             env_cfg.seed = agent_cfg["params"]["seed"]
 
             config_name = agent_cfg["params"]["config"]["name"]
diff --git a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_train_rl_games.py b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_train_rl_games.py
index 03b0dfadf2df..0aa892748095 100644
--- a/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_train_rl_games.py
+++ b/source/isaaclab/isaaclab/benchmark/entrypoints/backends/rl_games/benchmark_train_rl_games.py
@@ -194,7 +194,7 @@ def run(argv: list[str]) -> BenchmarkResult | None:
 
             apply_env_overrides(args_cli, env_cfg)
 
-            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli.device, seed=args_cli.seed)
+            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli)
 
             if args_cli.max_iterations is not None:
                 agent_cfg["params"]["config"]["max_epochs"] = args_cli.max_iterations
diff --git a/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.added.rst b/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.added.rst
new file mode 100644
index 000000000000..5eb347d43b69
--- /dev/null
+++ b/source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.added.rst
@@ -0,0 +1 @@
+* Added ``--seed`` to the RL-Games export command, matching the RSL-RL export command.
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rl_games.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rl_games.py
index 373368be2746..56cd204318fe 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rl_games.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rl_games.py
@@ -7,23 +7,24 @@
 
 from __future__ import annotations
 
+import argparse
+
 from ..common import resolve_seed
 
 
-def update_rl_games_cfg(agent_cfg: dict, device: str, *, seed: int | None = None) -> dict:
+def update_rl_games_cfg(agent_cfg: dict, args_cli: argparse.Namespace) -> dict:
     """Override an RL-Games agent configuration with the command-line arguments.
 
     Args:
         agent_cfg: The configuration for RL-Games agent.
-        device: Device the agent runs on, the simulation's device once :func:`~isaaclab.app.launch_simulation`
-            has resolved it.
-        seed: Seed override; ``-1`` draws a random seed.
+        args_cli: The command line arguments.
 
     Returns:
         The updated RL-Games agent configuration.
     """
     config = agent_cfg["params"]["config"]
-    config["device"] = config["device_name"] = device
-    if seed is not None:
-        agent_cfg["params"]["seed"] = resolve_seed(seed)
+    config["device"] = config["device_name"] = args_cli.device
+    if args_cli.seed is not None:
+        args_cli.seed = resolve_seed(args_cli.seed)
+        agent_cfg["params"]["seed"] = args_cli.seed
     return agent_cfg
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rl_games.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rl_games.py
index 1d9fe4faa208..0c4b1f60ab94 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rl_games.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rl_games.py
@@ -51,6 +51,7 @@ def parse_export_args(argv: list[str] | None = None) -> tuple[argparse.Namespace
     """Parse export arguments and return the remaining Hydra overrides."""
     parser = argparse.ArgumentParser(description="Export an RL agent with RL-Games.")
     add_common_export_args(parser, agent_default="rl_games_cfg_entry_point")
+    parser.add_argument("--seed", type=int, default=None, help="Seed used for the environment.")
     parser.add_argument(
         "--use_last_checkpoint",
         action="store_true",
@@ -106,7 +107,7 @@ def export_rl_games_agent(args_cli: argparse.Namespace, env_cfg: Any, agent_cfg:
     # initialized the selected backend
     from leapp import annotate
 
-    agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli.device)
+    agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli)
     params = agent_cfg["params"]
     env_cfg.scene.num_envs = 1
     env_cfg.seed = params["seed"]
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rl_games.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rl_games.py
index 4cc727222518..b4376e0356e1 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rl_games.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rl_games.py
@@ -112,7 +112,7 @@ def run(argv: list[str]) -> None:
         with launch_simulation(env_cfg, args_cli), contextlib.ExitStack() as cleanup:
             show_run_summary(screen, args_cli, env_cfg, library="rl_games", action="play")
             apply_env_overrides(args_cli, env_cfg)
-            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli.device, seed=args_cli.seed)
+            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli)
             params = agent_cfg["params"]
             env_cfg.seed = params["seed"]
 
diff --git a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rl_games.py b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rl_games.py
index 325f39f4451e..883e5a4029b5 100644
--- a/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rl_games.py
+++ b/source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rl_games.py
@@ -114,7 +114,7 @@ def run(argv: list[str]) -> None:
             show_run_summary(screen, args_cli, env_cfg, library="rl_games", action="train")
             apply_env_overrides(args_cli, env_cfg)
 
-            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli.device, seed=args_cli.seed)
+            agent_cfg = cli_args.update_rl_games_cfg(agent_cfg, args_cli)
             params = agent_cfg["params"]
             config = params["config"]
             if args_cli.max_iterations is not None: