Document camera-free experiment benchmarking

[cvolkcvolk]

Summary

Add single-GPU throughput and OSMO speedup results for a camera-free Arena experiment.

Detailed description

• Add single-GPU parallel-environment throughput results and plot.
• Add OSMO distributed-Run speedup results and plot.
• Link the performance reference from the documentation landing page.

[greptile-apps]

Greptile Summary

This PR adds preliminary performance documentation for Arena's two scaling axes:

• Documents single-GPU rollout throughput as the number of parallel environments increases.
• Documents OSMO speedup when independent Runs execute concurrently across GPUs.
• Adds accessible SVG charts, benchmark methodology, limitations, hardware details, and tested revisions.
• Links both benchmarks from the documentation landing page.

Confidence Score: 5/5

The documentation-only PR appears safe to merge with no actionable correctness, integration, or security issues identified.

The new page is correctly connected to the Sphinx toctree, its references and image paths resolve, and its benchmark values and chart scales are internally consistent.

Important Files Changed

Filename	Overview
docs/pages/references/performance.rst	Adds internally consistent benchmark methodology, results, limitations, and revision references for parallel environments and distributed Runs.
docs/index.rst	Adds valid links and concise benchmark results to the landing page and registers the performance page in the references toctree.
docs/images/performance/parallel_environment_throughput.svg	Adds an accessible single-GPU throughput chart whose labels, geometry, and values agree with the documentation.
docs/images/performance/distributed_run_speedup.svg	Adds an accessible distributed speedup chart whose labels, geometry, and values agree with the documentation.

_{Reviews (1): Last reviewed commit: "Document Arena performance scaling" | Re-trigger Greptile}

[arena-review-bot]

🤖 Isaac Lab-Arena Review Bot

Summary

Adds a Performance and scaling reference page with a single-GPU parallel-environment sweep and an OSMO distributed-Run sweep, two hand-written SVG charts, and two supporting sentences on the landing page. The write-up is careful — the scope caveats are honest, and I checked the arithmetic in both tables, the derived speedups, the minute conversions, and the bar geometry in both SVGs against the numbers they label: all consistent. My concerns are about reproducibility of the published numbers rather than the numbers themselves.

Design, Boundaries & Scope

This page presents itself as a reproducible reference, but the artifacts it depends on are not in the repo. isaaclab_arena_environments/experiment_configs/perflab/ does not exist on main; both pinned Arena commits belong to open PRs (#1151, #1188). A reader on a released Arena cannot run either benchmark, and the permalinks depend on commits that may become unreachable once those PRs are squash-merged. Worth deciding whether this page lands after those PRs, or brings the configs with it.

Findings

🟡 Warning: docs/pages/references/performance.rst:157 — the linked "camera-free benchmark configuration" sets num_envs: 1 and num_steps: 10, with an in-file comment calling that a provisional smoke length. The documented 300 steps and the environment sweep come from CLI overrides in the sibling perflab/README.md. Link the README too, or state which values were overridden. (The OSMO config link is exact — no issue there.)

🟡 Warning: docs/pages/references/performance.rst:153 — the tested revisions point at unmerged branches; see the scope note above.

🔵 Improvement: docs/index.rst:448 — the benchmark sentence sits under a claim about "thousands of heterogeneous (object-level) environments", but the workload used one object in all 1,024 environments; the perflab README explicitly says not to read the Rubik's-cube baseline as a heterogeneous-object result.

🔵 Improvement: docs/index.rst:480 — "active Arena execution" is undefined on the landing page, and with the OSMO overheads excluded the ~8x follows almost arithmetically from eight independent Runs on eight GPUs. The informative result is the sub-2% change in per-Run duration.

🔵 Improvement: docs/pages/references/performance.rst:53 — two-decimal throughput figures from single, unrepeated measurements overstate precision; three significant figures would match the stated methodology.

Test Coverage

Documentation only — no code paths, so no unit or simulation tests apply. The reproducibility gap noted above is the equivalent concern here: the benchmark configs and the PerfLab README that define the procedure should be in the repo for these numbers to be checkable.

Verdict

Minor fixes needed

[sangeetas-nv]

Looks beautiful and comprehensive. Thank you @cvolkcvolk !

[qianl-nv]

LGTM!