Cherrypick envs throughput benchmarking results

[xyao-nv]

Summary

Cherrypick envs throughput benchmarking results #1234

[greptile-apps]

Confidence Score: 4/5

The PR appears safe to merge after improving the non-blocking single-GPU benchmark reproduction reference.

Findings

1.  Benchmark Configuration Mismatch ▶

Summary

• Documents measured throughput across 1–1,024 parallel environments.
• Documents OSMO scaling across 1, 2, 4, and 8 concurrent GPUs.
• Adds benchmark scope, hardware details, tested revisions, and navigation links.
• The single-GPU reproduction link points to a provisional configuration that does not match the published run parameters.

_{Reviews (1) · Last reviewed commit: "Refine performance scaling wording"}

[arena-review-bot]

🤖 Isaac Lab-Arena Review Bot

Summary

Docs-only backport of #1234 (merged to main earlier today) onto release/0.3.0: a Performance and scaling reference page, two SVG charts, and two sentences on the landing page. I diffed all four files against main — they are byte-identical, so the cherry-pick is faithful and the only index.rst difference between the branches is the pre-existing development-version note. I re-checked the arithmetic independently (throughputs against the mean step times, the 2.02x/4.05x/7.95x speedups, the 20:55 → 2:38 conversion, 32 Runs, the sub-2% per-Run spread): all consistent.

Design, Boundaries & Scope

The one thing this backport changes versus the original is the permanence of the reproducibility gap. Tested revisions pins b0cd0b38e and 4ee056866, which exist only on the still-open PRs #1151 and #1188 — not on main, not on release/0.3.0 — and experiment_configs/perflab/ is on neither branch. On main that fixes itself when those PRs land; release/0.3.0 only takes cherry-picks, so a 0.3.0 reader will never be able to reproduce these numbers from the released tree, and the permalinks break if either branch is squash-merged and deleted. Worth deciding whether the configs should ride along with this backport.

Findings

🟡 Warning: docs/pages/references/performance.rst:151 — pinned revisions and benchmark configs are unreachable from release/0.3.0; see above.

🔵 Improvement: docs/pages/references/performance.rst:21 — "collected for this release" reads as 0.3.0 on this branch, but the measured revision has diverged from it.

The findings from the original review on #1234 carry over unchanged, since the content is identical — most relevantly that the linked camera_free_benchmark_experiment.yaml still sets num_envs: 1 / num_steps: 10 rather than the documented 300-step sweep, and that the landing-page sentence at docs/index.rst:443 sits directly under the "thousands of heterogeneous (object-level) environments" claim while the benchmark put the same rubiks_cube_hot3d_robolab in all 1,024 environments. Not re-posting them inline here; they were visible when #1234 merged.

Test Coverage

Documentation only — no code paths, so no unit or simulation tests apply. Docs build and pre-commit both pass on this branch.

Verdict

Minor fixes needed