Skip to content

Reduce lossless capture overhead and measure latency limits - #41

Draft
Tarzelf wants to merge 2 commits into
feat/godot-capture-performance-v01from
feat/godot-capture-latency-v02
Draft

Tarzelf wants to merge 2 commits into
feat/godot-capture-performance-v01from
feat/godot-capture-latency-v02

Conversation

@Tarzelf

@Tarzelf Tarzelf commented Nov 11, 2026 •

Copy link
Copy Markdown
Contributor

Rendering and durable recording hold up the local PiPER demonstration loop. This change uses faster lossless PNG compression, avoids redundant copies of newly allocated renderer arrays, reuses one independently copied context per step, and hashes/parses each bound document from the same read. Both native render passes, every physics tick, camera cadence, fsync barrier and recovery check remain in place.

The fixed 12-capture study preserves exact decoded RGB, NPY depth, native traces, commands, perception and learning arrays. Finalization averaged 19.4% shorter and files grew 2.70%. Total control timing is mixed: the candidate averaged 1.34% slower, including one pair 22.14% slower. About half the intervals still miss 20 ms. The single windowed pair improved from 14.175 to 13.555 seconds for five simulated seconds, without establishing a real-time guarantee.

Validation: 138 local test executions passed; Godot checks passed 7 + 29 assertions; four confirmation/windowed comparisons plus one diagnostic comparison passed; independent replay matched 510,000 values exactly; fresh outside-checkout base installation inspected/exported/reloaded 500 rows. The initial sparse-checkout fixture omission and its successful rerun are retained. Native graphics are tested on this Mac; no hardware or new learned-skill claim.

CI exposed an older exporter that intentionally accepts only frozen source hashes. The follow-up preserves its original pins, admits only the exact reviewed capture revision with a distinct source identity, and still rejects historical caller pins and unknown changes. All 65 reference/cache/preflight checks passed locally, and the updated helper was verified in an outside-checkout installation. Measured runtime files and original timing receipts are unchanged.

Stacked on #40. All attempts remain retained locally; compact receipts and reproduction scripts are included. No raw account footage, model weights or credentials. This is a draft research candidate, not a Hugging Face or research-site deployment.

Full report, frozen protocol and results

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant