[Test] Stabilize RTX MDL shader-warmup flake in rendering correctness tests

source/isaaclab_tasks/changelog.d/jichuanh-stabilize-rendering-mdl-warmup.rst

@@ -0,0 +1,13 @@
+Fixed
+^^^^^
+
+* Stabilized RTX MDL shader-warmup flakes in the rendering-correctness tests
+  by stepping the env 10 frames before reading the camera tensor, instead of
+  a single step. Affected ``test_shadow_hand_vision_presets.py``'s
+  ``test_camera_renders_not_empty`` and the three helper-driven tests in
+  ``rendering_test_utils.py``
+  (``rendering_test_shadow_hand``/``_cartpole``/``_dexsuite_kuka``) — all of
+  which intermittently failed with "Camera output is all zeros or all inf"
+  for ``simple_shading_*_mdl`` and ``simple_shading_constant_diffuse``
+  variants on cold-cache CI runners (the GPU returned a still-zero
+  framebuffer because the MDL material hadn't finished compiling).

source/isaaclab_tasks/test/rendering_test_utils.py

@@ -675,6 +675,24 @@ def validate_camera_outputs(
         pytest.fail(reason)
 
 
+def _warmup_render(env, num_steps: int = 10) -> None:
+    """Step the env ``num_steps`` times to let RTX MDL shaders finish compiling.
+
+    RTX MDL materials compile asynchronously on first use. A single ``env.step()``
+    is not enough — the GPU may return a still-zero framebuffer for shader
+    variants that have not finished compiling, which trips
+    :func:`validate_camera_outputs` with "no non-zero pixels" or
+    "all zeros or all inf". ``10`` is empirically enough across the MDL
+    presets (``simple_shading_constant_diffuse``, ``simple_shading_diffuse_mdl``,
+    ``simple_shading_full_mdl``) that flake on the CI runners as of 2026-05.
+    Adds ~1-2 s of wall time per parametrize variant; cheap relative to the
+    cost of a CI re-run.
+    """
+    actions = torch.zeros(env.num_envs, env.action_space.shape[-1], device=env.device)
+    for _ in range(num_steps):
+        env.step(actions)
+
+
 def rendering_test_shadow_hand(
     physics_backend: str,
     renderer: str,
@@ -700,6 +718,7 @@ def rendering_test_shadow_hand(
     try:
         env = ShadowHandVisionEnv(env_cfg)
         maybe_save_stage("shadow_hand", physics_backend, renderer, data_type)
+        _warmup_render(env)
 
         validate_camera_outputs(
             "shadow_hand",
@@ -739,6 +758,7 @@ def rendering_test_cartpole(
     try:
         env = CartpoleCameraEnv(env_cfg)
         maybe_save_stage("cartpole", physics_backend, renderer, data_type)
+        _warmup_render(env)
         validate_camera_outputs(
             "cartpole",
             physics_backend,
@@ -795,6 +815,7 @@ def rendering_test_dexsuite_kuka(
     try:
         env = ManagerBasedRLEnv(env_cfg)
         maybe_save_stage("dexsuite_kuka", physics_backend, renderer, data_type)
+        _warmup_render(env)
         validate_camera_outputs(
             "dexsuite_kuka",
             physics_backend,

source/isaaclab_tasks/test/test_shadow_hand_vision_presets.py

@@ -398,7 +398,12 @@ def render_correctness_env(request, shadow_hand_vision_presets):
     env = ShadowHandVisionEnv(cfg)
     env.reset()
     actions = torch.zeros(cfg.scene.num_envs, env.action_space.shape[-1], device=env.device)
-    env.step(actions)
+    # Step enough frames for RTX MDL shader variants to finish compiling. A single step
+    # returns a still-zero framebuffer for ``simple_shading_*_mdl`` presets on cold caches,
+    # which trips the assertion in ``test_camera_renders_not_empty`` below. 10 steps is
+    # empirically enough across the MDL variants that flake in CI as of 2026-05.
+    for _ in range(10):
+        env.step(actions)
     yield renderer_preset, camera_preset, physics, env
     env.close()