Skip to content

Reuse calibrated camera geometry and measure capture performance - #40

Draft
Tarzelf wants to merge 1 commit into
feat/godot-demonstration-v02from
feat/godot-capture-performance-v01
Draft

Tarzelf wants to merge 1 commit into
feat/godot-demonstration-v02from
feat/godot-capture-performance-v01

Conversation

@Tarzelf

@Tarzelf Tarzelf commented Nov 11, 2026

Copy link
Copy Markdown
Contributor

Repeated RGB-D frames rebuilt the same calibrated camera ray geometry. This change caches only those immutable directions, with exact calibration keys and a two-entry bound. Every frame still supplies fresh observations and runs the existing visibility, occlusion and missingness checks.

Stacked on #39 (feat/godot-demonstration-v02), whose 36 CI checks now pass.

The frozen comparison preserved exact observations, issued/applied actions, physics, task outcomes and learning exports. Average control-loop time fell from 6.152 s to 5.894 s across three paired two-second captures (4.2%), but one pair was 1.34% slower. The single windowed comparison also had a worse RPC p95. This is a modest preliminary optimization, not a reliable speedup or real-time qualification.

  • Cache retains at most 3.516 MiB for two 320×240 entries, or 14.063 MiB at the supported maximum resolution, excluding temporary allocations. No prior observations or object estimates are cached.
  • Adds separate profiling and exact-parity tools, registered budgets/source pins and every timing outcome. Physics remains 1 ms, controls 50 Hz, RGB-D 25 Hz; durability barriers are preserved.
  • 114 local test executions passed. Three confirmation pairs matched 306,000 native values, 306 RGB/depth files and 303 perception records, plus feature/target arrays, lifecycle, groups, configuration and data types. The windowed pair matched another 102,000 native values; a separate replay independently matched one candidate's 102,000 values and task score.
  • Fresh outside-checkout base installation passed CLI, inspection and 100-row export/reload checks. Native rendering was tested in the existing Mac simulation environment.

All eleven captures remain retained locally. No human demonstrations, fitting, hardware commands or new learned skill. The original 5/64 placement result remains unqualified. Diagnostic thread settings, cProfile overhead, shared-host variation and the slower measurements are disclosed in the report. This publishes source and compact evidence only; it is not a Hugging Face or research-site deployment.

Report, full results and reproduction

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