add settings to make training results deterministic

CONTRIBUTORS.md

@@ -183,6 +183,7 @@ Guidelines for modifications:
 * Xiaodi Yuan
 * Xinjie Yao
 * Xinpeng Liu
+* Xu Li
 * Yang Jin
 * Yanzi Zhu
 * Yijie Guo

apps/isaaclab.python.headless.determinism.kit

@@ -0,0 +1,171 @@
+##
+# Adapted from: https://github.com/NVIDIA-Omniverse/OmniIsaacGymEnvs/blob/main/apps/omni.isaac.sim.python.gym.camera.kit
+#
+# This app file designed specifically towards vision-based RL tasks. It provides necessary settings to enable
+# multiple cameras to be rendered each frame. Additional settings are also applied to increase performance when
+# rendering cameras across multiple environments.
+##
+
+[package]
+title = "Isaac Lab Python Headless Deterministic Rendering"
+description = "An app for running Isaac Lab headlessly with rendering enabled and deterministic settings"
+version = "3.0.0"
+
+# That makes it browsable in UI with "experience" filter
+keywords = ["experience", "app", "isaaclab", "python", "camera", "minimal"]
+
+[dependencies]
+# Isaac Lab minimal app
+"isaaclab.python.headless" = {}
+"omni.replicator.core" = {}
+
+# Rendering
+"omni.kit.material.library" = {}
+"omni.kit.viewport.rtx" = {}
+"omni.kit.hydra_texture" = {}  # required for headless rendering and rtx instantiation
+
+
+[settings.isaaclab]
+# This is used to check that this experience file is loaded when using cameras
+cameras_enabled = true
+
+[settings]
+# Note: This path was adapted to be respective to the kit-exe file location
+app.versionFile = "${exe-path}/VERSION"
+app.folder = "${exe-path}/"
+app.name = "IsaacLab"
+app.version = "3.0.0"
+
+### FSD
+app.useFabricSceneDelegate = true
+# Temporary, should be enabled by default in Kit soon
+rtx.hydra.readTransformsFromFabricInRenderDelegate = true
+
+# Disable print outs on extension startup information
+# this only disables the app print_and_log function
+app.enableStdoutOutput = false
+
+# set the default ros bridge to disable on startup
+isaac.startup.ros_bridge_extension = ""
+
+# Flags for better rendering performance
+# Disabling these settings reduces renderer VRAM usage and improves rendering performance, but at some quality cost
+rtx.translucency.enabled = false
+rtx.reflections.enabled = false
+rtx.indirectDiffuse.enabled = false
+rtx-transient.dlssg.enabled = false
+rtx.directLighting.sampledLighting.enabled = true
+rtx.directLighting.sampledLighting.samplesPerPixel = 1
+rtx.sceneDb.ambientLightIntensity = 1.0
+# rtx.shadows.enabled = false
+
+# Avoids replicator warning
+rtx.pathtracing.maxSamplesPerLaunch = 1000000
+# Avoids silent trimming of tiles
+rtx.viewTile.limit = 1000000
+
+# Disable present thread to improve performance
+exts."omni.renderer.core".present.enabled=false
+
+# Disabling these settings reduces renderer VRAM usage and improves rendering performance, but at some quality cost
+rtx.raytracing.cached.enabled = false
+rtx.ambientOcclusion.enabled = false
+
+# Set the DLSS model
+rtx.post.dlss.execMode = 0 # can be 0 (Performance), 1 (Balanced), 2 (Quality), or 3 (Auto)
+
+# Avoids unnecessary GPU context initialization
+renderer.multiGpu.maxGpuCount=1
+
+# Force synchronous rendering to improve training results
+omni.replicator.asyncRendering = false
+
+# Avoids frame offset issue
+app.updateOrder.checkForHydraRenderComplete = 1000
+app.renderer.waitIdle=true
+app.hydraEngine.waitIdle=true
+
+# NuRec
+omni.rtx.nre.compositing.rendererHints = 3
+
+# Forces serial processing for omni graph to avoid NCCL timeout hangs in distributed training
+app.execution.debug.forceSerial = true
+
+app.audio.enabled = false
+
+# Enable Vulkan - avoids torch+cu12 error on windows
+app.vulkan = true
+
+# Set profiler backend to NVTX by default
+app.profilerBackend = "nvtx"
+
+# Disables rate limit in runloop
+app.runLoops.main.rateLimitEnabled=false
+
+# disable replicator orchestrator for better runtime perf
+exts."omni.replicator.core".Orchestrator.enabled = false
+
+# disable the metrics assembler change listener, we don't want to do any runtime changes
+metricsAssembler.changeListenerEnabled = false
+
+# explicitly disable omni.kit.pip_archive to prevent conflicting dependencies
+app.extensions.excluded = ["omni.kit.pip_archive", "omni.isaac.ml_archive", "isaacsim.pip.newton", "omni.warp.core"]
+
+[settings.app.python]
+# These disable the kit app from also printing out python output, which gets confusing
+interceptSysStdOutput = false
+logSysStdOutput = false
+
+[settings.app.renderer]
+skipWhileMinimized = false
+sleepMsOnFocus = 0
+sleepMsOutOfFocus = 0
+
+[settings.physics]
+updateToUsd = false
+updateParticlesToUsd = false
+updateVelocitiesToUsd = false
+updateForceSensorsToUsd = false
+outputVelocitiesLocalSpace = false
+useFastCache = false
+visualizationDisplayJoints = false
+fabricUpdateTransformations = false
+fabricUpdateVelocities = false
+fabricUpdateForceSensors = false
+fabricUpdateJointStates = false
+### When Direct GPU mode is enabled (suppressReadback=true) use direct interop between PhysX GPU and Fabric
+fabricUseGPUInterop = true
+
+# Register extension folder from this repo in kit
+[settings.app.exts]
+folders = [
+    "${exe-path}/exts",  # kit extensions
+    "${exe-path}/extscore",  # kit core extensions
+    "${exe-path}/../exts",  # isaac extensions
+    "${exe-path}/../extscache",  # isaac cache extensions
+    "${exe-path}/../extsPhysics",  # isaac physics extensions
+    "${exe-path}/../isaacsim/exts",  # isaac extensions for pip
+    "${exe-path}/../isaacsim/extscache",  # isaac cache extensions for pip
+    "${exe-path}/../isaacsim/extsPhysics",  # isaac physics extensions for pip
+    "${app}", # needed to find other app files
+    "${app}/../source", # needed to find extensions in Isaac Lab
+]
+
+[settings.persistent]
+UJITSO.geometry = true
+UJITSO.enabled = true
+
+# Asset path
+# set the S3 directory manually to the latest published S3
+# note: this is done to ensure prior versions of Isaac Sim still use the latest assets
+[settings]
+persistent.isaac.asset_root.default = "https://omniverse-content-staging.s3-us-west-2.amazonaws.com/Assets/Isaac/6.0"
+persistent.isaac.asset_root.cloud = "https://omniverse-content-staging.s3-us-west-2.amazonaws.com/Assets/Isaac/6.0"
+persistent.isaac.asset_root.nvidia = "https://omniverse-content-staging.s3-us-west-2.amazonaws.com/Assets/Isaac/6.0"
+
+rtx.rendermode = "RealTimePathTracing"
+# Determinism: disable RTPT caches so each run recomputes lighting from the same
+# explicit inputs. Warm/cold cache history can otherwise change accumulation paths
+# and produce frame-to-frame drift across repeated experiments.
+rtx.rtpt.cached.enabled = false
+rtx.rtpt.lightcache.cached.enabled = false

docs/source/features/reproducibility.rst

@@ -20,6 +20,33 @@ simulation results are reproducible across different runs. The seed is set into
 parameters :attr:`isaaclab.envs.ManagerBasedEnvCfg.seed` or :attr:`isaaclab.envs.DirectRLEnvCfg.seed`
 depending on the manager-based or direct environment implementation respectively.
 
+.. note::
+
+   **Strict PyTorch determinism is only wired into the RL-Games training script.**
+
+   The optional ``--deterministic`` flag (which calls :meth:`~isaaclab.utils.seed.configure_seed` with
+   ``torch_deterministic=True``, enabling ``torch.use_deterministic_algorithms(True)`` and related CUDNN
+   settings) exists only on ``scripts/reinforcement_learning/rl_games/train.py``.
+
+   The RSL-RL, Stable-Baselines3, SKRL, and RLinf training scripts under
+   ``scripts/reinforcement_learning/`` still honor ``--seed`` / agent configuration for the Isaac Lab
+   environment and learning stack, but they do **not** expose an equivalent opt-in for strict PyTorch-wide
+   deterministic algorithms. If you need that behavior with another framework, call
+   :meth:`~isaaclab.utils.seed.configure_seed` with ``torch_deterministic=True`` from your own training
+   entry point, keeping in mind that some CUDA operations may error when no deterministic implementation
+   exists.
+
+To enable deterministic rendering/app settings, launch workflows with the deterministic experience file:
+
+.. code-block:: bash
+
+  ./isaaclab.sh -p scripts/reinforcement_learning/rl_games/train.py \
+    --task Isaac-Cartpole-v0 \
+    --experience isaaclab.python.headless.determinism.kit
+
+For RL-Games, combine this with ``--deterministic`` if you also want strict PyTorch deterministic
+algorithms in addition to deterministic app/render settings.
+
 For results on our determinacy testing for RL training, please check the GitHub Pull Request `#940`_.
 
 .. tip::

scripts/reinforcement_learning/rl_games/train.py

@@ -25,6 +25,7 @@
 from isaaclab.utils.assets import retrieve_file_path
 from isaaclab.utils.dict import print_dict
 from isaaclab.utils.io import dump_yaml
+from isaaclab.utils.seed import configure_seed
 
 from isaaclab_rl.rl_games import MultiObserver, PbtAlgoObserver, RlGamesGpuEnv, RlGamesVecEnvWrapper
 
@@ -48,6 +49,15 @@
     "--agent", type=str, default="rl_games_cfg_entry_point", help="Name of the RL agent configuration entry point."
 )
 parser.add_argument("--seed", type=int, default=None, help="Seed used for the environment")
+parser.add_argument(
+    "--deterministic",
+    action="store_true",
+    default=False,
+    help=(
+        "Enable strict PyTorch determinism (torch.use_deterministic_algorithms(True), cudnn deterministic, "
+        "etc.). Opt-in: some CUDA ops may error if no deterministic implementation exists."
+    ),
+)
 parser.add_argument(
     "--distributed", action="store_true", default=False, help="Run training with multiple GPUs or nodes."
 )
@@ -203,6 +213,8 @@ def main():
         else:
             runner = Runner(IsaacAlgoObserver())
 
+        configure_seed(env_cfg.seed, args_cli.deterministic)
+
         runner.load(agent_cfg)
         runner.reset()