Request determinism from the play entrypoints too

apply_env_overrides is a training-entrypoint helper, so --deterministic reached
neither the physics config nor Warp's global on the play paths, which build their
own environments. They now share a request_determinism() helper, called before the
environment exists so Warp's module-build-time read still sees it.

Also qualify the reproducibility docs: the flag configures Warp through the RL
entrypoints, not through AppLauncher, so a script that builds its own environment
must set warp.config.deterministic itself.

Author: hujc7

docs/source/features/reproducibility.rst

@@ -43,7 +43,9 @@ which the flag sets (see :ref:`below <reproducibility-warp-determinism>`).
 **Warp determinism.** Newton's solvers accept a ``deterministic`` argument that Isaac Lab supplies
 from :attr:`~isaaclab.physics.PhysicsCfg.deterministic`, and they apply it as a per-module option.
 Its sensor and geometry kernels take no such argument and fall back to ``warp.config.deterministic``,
-so ``--deterministic`` sets that global to ``RUN_TO_RUN`` as well. Without it the scene BVH is built
+so the RL training and play entrypoints set that global to ``RUN_TO_RUN`` when
+``--deterministic`` is passed. A script that builds its own environment without those entrypoints
+must set ``warp.config.deterministic`` itself, before the environment is created. Without it the scene BVH is built
 over an atomically compacted shape list whose order varies between processes, and a tiled camera can
 render a few pixels differently from identical simulation state. An explicitly chosen mode, such as
 ``GPU_TO_GPU``, is left untouched.

source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rl_games.py

@@ -30,6 +30,7 @@
     apply_video_recording,
     create_isaaclab_env,
     pre_launch_video_config,
+    request_determinism,
     resolve_checkpoint_selector,
     resolve_play_task_name,
     show_run_summary,
@@ -107,6 +108,8 @@ def main():
             train_task_name = task_name.replace("-Play", "")
 
             env_cfg.scene.num_envs = args_cli.num_envs if args_cli.num_envs is not None else env_cfg.scene.num_envs
+            # Warp reads its determinism mode at module build time, so request it before the env exists.
+            request_determinism(args_cli, env_cfg)
             env_cfg.sim.device = args_cli.device if args_cli.device is not None else env_cfg.sim.device
 
             if args_cli.seed == -1:

source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_rsl_rl.py

@@ -29,6 +29,7 @@
     apply_video_recording,
     create_isaaclab_env,
     pre_launch_video_config,
+    request_determinism,
     resolve_checkpoint_selector,
     resolve_play_task_name,
     show_run_summary,
@@ -126,6 +127,8 @@ def main(env_cfg: ManagerBasedRLEnvCfg | DirectRLEnvCfg | DirectMARLEnvCfg, agen
 
             agent_cfg = cli_args.update_rsl_rl_cfg(agent_cfg, args_cli)
             env_cfg.scene.num_envs = args_cli.num_envs if args_cli.num_envs is not None else env_cfg.scene.num_envs
+            # Warp reads its determinism mode at module build time, so request it before the env exists.
+            request_determinism(args_cli, env_cfg)
 
             agent_cfg = handle_deprecated_rsl_rl_cfg(agent_cfg, installed_version)
 

source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_sb3.py

@@ -27,6 +27,7 @@
     apply_video_recording,
     create_isaaclab_env,
     pre_launch_video_config,
+    request_determinism,
     resolve_checkpoint_selector,
     resolve_play_task_name,
     show_run_summary,
@@ -112,6 +113,8 @@ def main():
                 args_cli.seed = random.randint(0, 10000)
 
             env_cfg.scene.num_envs = args_cli.num_envs if args_cli.num_envs is not None else env_cfg.scene.num_envs
+            # Warp reads its determinism mode at module build time, so request it before the env exists.
+            request_determinism(args_cli, env_cfg)
             agent_cfg["seed"] = args_cli.seed if args_cli.seed is not None else agent_cfg["seed"]
             env_cfg.seed = agent_cfg["seed"]
             env_cfg.sim.device = args_cli.device if args_cli.device is not None else env_cfg.sim.device

source/isaaclab_rl/isaaclab_rl/entrypoints/backends/play_skrl.py

@@ -32,6 +32,7 @@
     create_isaaclab_env,
     pre_launch_video_config,
     preserve_attribute,
+    request_determinism,
     resolve_checkpoint_selector,
     resolve_play_task_name,
     show_run_summary,
@@ -165,6 +166,8 @@ def _main():
             train_task_name = task_name.replace("-Play", "")
 
             env_cfg.scene.num_envs = args_cli.num_envs if args_cli.num_envs is not None else env_cfg.scene.num_envs
+            # Warp reads its determinism mode at module build time, so request it before the env exists.
+            request_determinism(args_cli, env_cfg)
             env_cfg.sim.device = args_cli.device if args_cli.device is not None else env_cfg.sim.device
 
             # configure the ML framework into the global skrl variable

source/isaaclab_rl/isaaclab_rl/entrypoints/common.py

@@ -493,11 +493,25 @@ def apply_env_overrides(args_cli: argparse.Namespace, env_cfg: Any, *, apply_dev
     # --deterministic is an AppLauncher flag, so it only reaches carb settings on its own.
     # Record the request on the resolved physics config; each backend translates and validates
     # it when the simulation starts.
-    if getattr(args_cli, "deterministic", False):
-        physics_cfg = getattr(getattr(env_cfg, "sim", None), "physics", None)
-        if physics_cfg is not None:
-            physics_cfg.deterministic = True
-        request_warp_determinism(physics_cfg)
+    request_determinism(args_cli, env_cfg)
+
+
+def request_determinism(args_cli: argparse.Namespace, env_cfg: Any) -> None:
+    """Record a ``--deterministic`` request on the config tree and on Warp's global.
+
+    Call this before the environment is created: Warp reads its setting at module build time
+    and the scene BVH is built while the environment is constructed.
+
+    Args:
+        args_cli: Parsed command-line arguments.
+        env_cfg: Isaac Lab environment config.
+    """
+    if not getattr(args_cli, "deterministic", False):
+        return
+    physics_cfg = getattr(getattr(env_cfg, "sim", None), "physics", None)
+    if physics_cfg is not None:
+        physics_cfg.deterministic = True
+    request_warp_determinism(physics_cfg)
 
 
 def request_warp_determinism(physics_cfg: Any) -> None: