Add opt-in async OVRTX rendering

[pv-nvidia]

Description

Adds an opt-in asynchronous render path to the OVRTX renderer. Rendering can overlap with simulation and Python work, improving throughput. It is off by default (async_rendering=False), so existing behavior is unchanged.

What this adds

A single on/off switch, OVRTXRendererCfg.async_rendering (default False):

• False: synchronous rendering (unchanged blocking behavior).
• True: asynchronous rendering. render() submits the render and returns immediately, so rendering overlaps with simulation and Python work; camera outputs arrive one frame later. The render pipeline depth is fixed at two (one frame of camera latency); exposing per-frame latency control is deferred to a future Isaac Lab-side setting rather than a renderer-level knob.

Internally the renderer delegates frame execution to a small _RenderStrategy hierarchy: _SyncRenderStrategy writes transforms straight into OVRTX and consumes each step inline, while _AsyncRenderStrategy pipelines steps and double-buffers transform staging. The async object is created only when async is enabled.

For tests, the OVRTX_ASYNC_RENDERING environment variable overrides async_rendering (0/false/off disable, any other non-empty value enables).

The runtime benchmark now supports untimed warmup frames and a boolean --ovrtx_async_rendering flag, so sync/async camera runs can use scripts/benchmarks/runtime.py directly.

CI support also ensures a per-commit ECR image tag exists when the dependency cache is reused locally, so downstream package-test jobs can pull the image across self-hosted runners.

Correctness

The RGB Cartpole golden-image test passes on the async path (test_rendering_cartpole_kitless.py::test_rendering_cartpole_kitless[newton-ovrtx-rgb]). Async output matches the sync path (1 frame apart) within the test's SSIM / pixel-difference tolerances.

The OVRTX clone-plan unit tests also pass after the async-state refactor (source/isaaclab_ov/test/test_ovrtx_clone_plan.py, 14 tests), including the backend-free _initialize_from_spec path used for combined-stage dump coverage.

Benchmarks

Measured with the official runtime benchmark (scripts/benchmarks/runtime.py) using kitless OVRTX RGB camera tasks, --warmup_frames 30, and --num_frames 200. SPS = environment steps/sec (num_envs / mean_step_time_s). These are single local runs with schema output, so treat them as throughput signals rather than a multi-repetition benchmark suite. "async" is the one-frame-latency asynchronous path.

Benchmark provenance for the runs below:

• Git commit: 318828f57e0f9b85c760caeae3630c79b3e604ce
• OVRTX: 0.3.0.312915
• Selectors: physics=newton_mjwarp renderer=ovrtx presets=rgb

Cartpole camera

Task: Isaac-Cartpole-Camera-Direct

envs	sync mean ms	sync SPS	async mean ms	async SPS	speedup
64	14.281	4,481.4	9.761	6,556.7	1.46x
256	18.733	13,665.6	13.279	19,277.9	1.41x
1024	31.460	32,549.0	21.972	46,605.3	1.43x

Command used for each row (add --ovrtx_async_rendering for the async column, --no-ovrtx_async_rendering for sync):

python scripts/benchmarks/runtime.py \
    --task Isaac-Cartpole-Camera-Direct \
    --num_envs <64|256|1024> \
    --num_frames 200 \
    --warmup_frames 30 \
    --benchmark_formatter schema \
    --output_path ./results/cartpole_<envs>_<sync_or_async> \
    --headless \
    --ovrtx_async_rendering \
    physics=newton_mjwarp renderer=ovrtx presets=rgb

Shadow Hand camera benchmark

Task: Isaac-Reorient-Cube-Shadow-Camera-Benchmark-Direct

envs	sync mean ms	sync SPS	async mean ms	async SPS	speedup
64	22.601	2,831.7	19.294	3,317.1	1.17x
256	30.151	8,490.5	24.017	10,659.3	1.26x
1024	53.297	19,212.9	36.996	27,679.0	1.44x

Command used for each row (add --ovrtx_async_rendering for the async column, --no-ovrtx_async_rendering for sync):

python scripts/benchmarks/runtime.py \
    --task Isaac-Reorient-Cube-Shadow-Camera-Benchmark-Direct \
    --num_envs <64|256|1024> \
    --num_frames 200 \
    --warmup_frames 30 \
    --benchmark_formatter schema \
    --output_path ./results/shadow_hand_<envs>_<sync_or_async> \
    --headless \
    --ovrtx_async_rendering \
    physics=newton_mjwarp renderer=ovrtx presets=rgb

Enabling async adds one frame of camera latency.

Type of change

• New feature (non-breaking change which adds functionality)

Screenshots

Not applicable — performance/API change with no visual UI. Async RGB output is validated against the sync path via the golden correctness test above (SSIM ~0.9997, ~0% pixel diff).

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
• 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 (do not edit CHANGELOG.rst or bump extension.toml — CI handles that)
• I have added my name to the CONTRIBUTORS.md or my name already exists there

[pv-nvidia]

So this becomes self._async is not None

[r-schmitt]

do we want to raise a ValueError here? it seems like we could instead log a warning about the argument's compatibility instead and ignore the arg when its not applicable