Fix OVRTX GPU transform default

[ruziniuuuuu]

Description

Default OVRTX's environment-controlled read_gpu_transforms setting to the safe CPU-read path. OVRTX warns that enabling Fabric transform reads together with geometry streaming can prevent dynamic geometry from streaming correctly; in Newton workloads this intermittently appears as missing robot links in OVRTX camera output.

The existing ISAAC_LAB_OVRTX_READ_GPU_TRANSFORMS=1 override remains available as an explicit opt-in for workloads that have validated the previous GPU path. No dependency changes are required.

A standalone four-environment differential reproducer found:

Configuration	Missing-link reproductions
OVRTX 0.3, current GPU-transform default	2 / 3
OVRTX 0.3, GPU transform reads disabled	0 / 10
OVRTX 0.4, current GPU-transform default	1 / 5
OVRTX 0.4, GPU transform reads disabled	0 / 4 completed runs

The Newton viewer-disabled control reproduced 0 / 3 times. USD visibility stayed unchanged and Newton body/joint states stayed finite and synchronized during captured failures.

Fixes #6625

Type of change

• Bug fix (non-breaking change which fixes an issue)

Screenshots

Not applicable to the adapter default itself. Issue #6625 describes the captured missing-link evidence and standalone reproducer.

Validation

• Regression test was added first and failed against the previous default.
• pytest source/isaaclab_ov/test/test_ovrtx_renderer_contract.py -q: 14 passed.
• ./isaaclab.sh -p tools/changelog/cli.py check develop: passed.
• ./isaaclab.sh -f: passed before commit and again before push.
• Standalone MRE: OVRTX 0.3 safe path completed 10 / 10 runs without the symptom; OVRTX 0.4 safe path completed 4 / 4 runs without it.

Current-develop end-to-end validation used its exact Newton pin (newton 1.4.0.dev0, warp-lang 1.15.0.dev20260626) with OVRTX 0.3: the standalone Newton-viewer MRE completed 1,000 steps across four environments with reproduced=False, finite/synchronized state, and no unsafe transform/geometry-streaming warning. A separate RTX/MDL shutdown segfault occurred only after the result and final capture were written.

Risk

The CPU transform-read path may have different renderer performance characteristics. Explicit opt-in to the previous behavior is preserved with ISAAC_LAB_OVRTX_READ_GPU_TRANSFORMS=1.

Checklist

• I have read and understood the contribution guidelines
• I have run the pre-commit checks with ./isaaclab.sh --format
• I have made corresponding changes to the documentation (changelog fragment)
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• I have added a changelog fragment under source/<pkg>/changelog.d/ for every touched package
• I have added my name to the CONTRIBUTORS.md or my name already exists there

[ruziniuuuuu]

Standalone reproducer and captured output

The complete reproducer is published at an immutable commit:

• README and run instructions
• Minimal Python entry point (repro.py)
• Robot USD, payloads, meshes, materials, and textures
• Complete captured state and detector evidence

It can be run directly from that folder:

python repro.py --num-envs 4 --steps 5000 --seed 17

# Automated fresh-process trial
python repro.py --headless --steps 1000 --stop-on-failure

The repository contains the robot asset and does not import GalbotLab. Four environments receive identical robot and camera states; the detector compares persistent denoised silhouettes and only reports a rendering failure while Newton state is finite and synchronized. The entry point supports both the stable and current develop launcher locations.

Reference OVRTX frame (env_0)	Missing-link OVRTX frame (env_1)

The captured failure affected environment 1 with 399 differing silhouette pixels (352 persistent across the three-frame confirmation window, maximum RGB channel error 211). USD visibility remained unchanged; the exact camera poses, body/joint state, package versions, and detection values are included in evidence/.

Newton visualizer framebuffer from the same capture:

Fix validation on current develop

I overlaid the PR head (42ba436fec2f25cd7ccf47889f1ee555cce28f52) onto current Isaac Lab develop and used its exact Newton dependency (c7ae7c7648cd0717df39e5c94b95d5a02c997320, newton 1.4.0.dev0, warp-lang 1.15.0.dev20260626) with OVRTX 0.3.0.312915.

The Newton-viewer/OVRTX run completed 1,000 simulation and render steps across four environments with reproduced=False; all compared body/joint states stayed finite and synchronized. The unsafe readTransformsFromFabricInRenderDelegate plus geometry-streaming warning was absent. After the result and final capture were written, Isaac Sim hit a separate RTX/MDL shutdown segfault; it did not occur during the test loop and is not counted as validation of that cleanup path.

[greptile-apps]

Greptile Summary

This PR changes the default for OVRTX's read_gpu_transforms setting from enabled ("1") to disabled ("0"), fixing intermittent missing robot links in Newton/OVRTX camera output caused by a renderer-side incompatibility between Fabric transform reads and RTX geometry streaming. The previous GPU-read path remains available as an explicit opt-in via ISAAC_LAB_OVRTX_READ_GPU_TRANSFORMS=1.

• ovrtx_renderer.py: One-line default change in _read_gpu_transforms_enabled() with an explanatory comment; the function is consumed at OVRTXRenderer.__init__ time when constructing RendererConfig.
• test_ovrtx_renderer_contract.py: Two new contract tests verify the disabled default and the explicit opt-in path using monkeypatch against the env var.
• changelog.d/ and CONTRIBUTORS.md: Standard housekeeping fragments added.

Confidence Score: 5/5

Safe to merge — the change is a one-line default flip with a well-tested escape hatch preserved via env var.

The fix is narrowly scoped to a single default value in a helper function. The CPU read path was already a supported, tested code branch; only the default changes. Two new contract tests directly cover the new default and the opt-in path, and the existing test suite still passes. The env-var override ensures no workloads are permanently broken by the new default.

No files require special attention.

Important Files Changed

Filename	Overview
source/isaaclab_ov/isaaclab_ov/renderers/ovrtx_renderer.py	Single-line default change from "1" to "0" in _read_gpu_transforms_enabled(); well-commented, no structural changes.
source/isaaclab_ov/test/test_ovrtx_renderer_contract.py	Imports _read_gpu_transforms_enabled and adds two focused monkeypatch tests covering the new default and opt-in; properly guarded by the module-level skipif.
source/isaaclab_ov/changelog.d/codex-fix-newton-ovrtx-transforms.rst	New changelog fragment documenting the fix and the opt-in env var; content is accurate.
CONTRIBUTORS.md	Adds "Chenyu Cao" in alphabetical order; no issues.

Sequence Diagram

%%{init: {'theme': 'neutral'}}%%
sequenceDiagram
    participant Env as Environment Variable ISAAC_LAB_OVRTX_READ_GPU_TRANSFORMS
    participant Fn as _read_gpu_transforms_enabled()
    participant Init as OVRTXRenderer.__init__
    participant RC as RendererConfig
    participant R as ovrtx.Renderer

    Init->>Fn: call at construction time
    Fn->>Env: os.environ.get(key, "0")
    alt env var unset (default)
        Env-->>Fn: "0"
        Fn-->>Init: False (CPU path)
    else "env var = "1" (opt-in)"
        Env-->>Fn: "1"
        Fn-->>Init: True (GPU path)
    else env var invalid
        Fn-->>Init: raises ValueError
    end
    Init->>RC: "RendererConfig(read_gpu_transforms=False/True)"
    RC->>R: Renderer(config)

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
sequenceDiagram
    participant Env as Environment Variable ISAAC_LAB_OVRTX_READ_GPU_TRANSFORMS
    participant Fn as _read_gpu_transforms_enabled()
    participant Init as OVRTXRenderer.__init__
    participant RC as RendererConfig
    participant R as ovrtx.Renderer

    Init->>Fn: call at construction time
    Fn->>Env: os.environ.get(key, "0")
    alt env var unset (default)
        Env-->>Fn: "0"
        Fn-->>Init: False (CPU path)
    else "env var = "1" (opt-in)"
        Env-->>Fn: "1"
        Fn-->>Init: True (GPU path)
    else env var invalid
        Fn-->>Init: raises ValueError
    end
    Init->>RC: "RendererConfig(read_gpu_transforms=False/True)"
    RC->>R: Renderer(config)

Loading

_{Reviews (1): Last reviewed commit: "Fix OVRTX GPU transform default" | Re-trigger Greptile}

[pbarejko]

Hi, thank you for the PR.

I wouldn't merge this change. Unfortunately, I am sad to say, this doesn't fix anything. The problem is upstream - Fabric. This PR hides the symptom and puts Isaac Lab on very slow code path - it would have huge performance implications. My best recommendation is to use ISAAC_LAB_OVRTX_READ_GPU_TRANSFORMS=0 in your local workflow.

This is a known issue that has been reported upstream and we should have fix for this.

@mataylor-nvidia can you check:

1. Rayyan's fix in Fabric is available in ovrtx-0.4 release? If it is this would confirm if there another hiding bug somewhere and we would have to file another nvbug with provided repro steps.
2. If it's not, I know you have built Rayyan's branch, can you use it to validate that it fixes @ruziniuuuuu problem without flipping read_gpu_transforms?

[ruziniuuuuu]

I retested ISAAC_LAB_OVRTX_READ_GPU_TRANSFORMS=0 on Isaac Lab v3.0.0-beta2.patch1. Setting it alone does not work because this version does not read the variable and still passes read_gpu_transforms=True to OVRTX.

Environment variable only: links still disappear

env_0	env_2

I then added a small local backport so that 0 actually reaches RendererConfig(read_gpu_transforms=False).

Environment variable with backport: links remain complete

env_0	env_2

With the backport, 3 independent 1,000-step runs completed with reproduced=False, and the Fabric/geometry-streaming warning disappeared.

Updated standalone reproducer and evidence

[pbarejko]

I retested ISAAC_LAB_OVRTX_READ_GPU_TRANSFORMS=0 on Isaac Lab v3.0.0-beta2.patch1. Setting it alone does not work because this version does not read the variable and still passes read_gpu_transforms=True to OVRTX.

Environment variable only: links still disappear

env_0 env_2

I then added a small local backport so that 0 actually reaches RendererConfig(read_gpu_transforms=False).

Environment variable with backport: links remain complete

env_0 env_2

With the backport, 3 independent 1,000-step runs completed with reproduced=False, and the Fabric/geometry-streaming warning disappeared.

Updated standalone reproducer and evidence

Good you have a workaround! In long term we will fix the problem upstream in OvRTX and you wouldn't have to manipulate read-gpu-transforms variable. This problem should be addressed in next release of ovrtx-0.5.

[pbarejko]

Hi @ruziniuuuuu, we have verified upstream fix for this issue: #6625 (comment)

Therefore I am taking the liberty and closing this PR since this is not applicable solution.