[Mac OS] [RL] Run RL agents on the simulation device and raise without CUDA

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 set up their agent configuration with
+  the same shared overrides as the training and play workflows, including the device the simulation runs on.

source/isaaclab/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.fixed.rst

@@ -0,0 +1,2 @@
+* 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``.

source/isaaclab/isaaclab/app/sim_launcher.py

@@ -504,10 +504,15 @@ def _resolve_device(sim_cfg, args: dict, launchers: list[SimulationLauncher]) ->
 
     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.
+
+    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():
+        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)}"

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"]

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
@@ -194,19 +194,15 @@ 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"]
+            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
 
             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)

source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_play_rsl_rl.py

@@ -139,6 +139,7 @@ def run(argv: list[str]) -> BenchmarkResult:
                 env_cfg.scene.num_envs = args.num_envs
             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))

source/isaaclab/isaaclab/benchmark/entrypoints/backends/rsl_rl/benchmark_train_rsl_rl.py

@@ -192,9 +192,8 @@ def run(argv: list[str]) -> BenchmarkResult | None:
             env_cfg.seed = agent_cfg.seed
 
             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)

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))
@@ -212,7 +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=agent_cfg["device"], print_system_info=True)
 
             def policy(obs):
                 """Map an observation batch to a deterministic action batch via the sb3 agent.

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
@@ -240,7 +241,7 @@ 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"]
 
             # Convert the iteration override to SB3 total timesteps.

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.

source/isaaclab_rl/changelog.d/antoiner-fix-rsl-rl-runner-cpu-fallback.changed.rst

@@ -0,0 +1,3 @@
+* **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.

source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rl_games.py

@@ -0,0 +1,30 @@
+# 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
+
+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.
+
+    Args:
+        agent_cfg: The configuration for RL-Games agent.
+        args_cli: The command line arguments.
+
+    Returns:
+        The updated RL-Games agent configuration.
+    """
+    config = agent_cfg["params"]["config"]
+    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

source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_rsl_rl.py

@@ -94,6 +94,7 @@ 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

source/isaaclab_rl/isaaclab_rl/entrypoints/backends/cli_args_sb3.py

@@ -0,0 +1,29 @@
+# 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.
+
+    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 args_cli.seed is not None:
+        args_cli.seed = resolve_seed(args_cli.seed)
+        agent_cfg["seed"] = args_cli.seed
+    return agent_cfg

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,
@@ -50,6 +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",
@@ -105,6 +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)
     params = agent_cfg["params"]
     env_cfg.scene.num_envs = 1
     env_cfg.seed = params["seed"]

source/isaaclab_rl/isaaclab_rl/entrypoints/backends/export_rsl_rl.py

@@ -137,6 +137,7 @@ def export_rsl_rl_agent(args_cli: argparse.Namespace, env_cfg: Any, agent_cfg: R
         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

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__)
 
@@ -112,10 +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"]
-            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"]))

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)))
@@ -147,7 +145,7 @@ 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=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)

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__)
 
@@ -114,18 +114,14 @@ 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
             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"]
 

source/isaaclab_rl/isaaclab_rl/entrypoints/backends/train_rsl_rl.py

@@ -126,7 +126,6 @@ def _run(args_cli: argparse.Namespace) -> None:
                 agent_cfg.max_iterations = args_cli.max_iterations
             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
 

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__)
 
@@ -115,9 +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)
-            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"]

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):