Slicecam fine-acquire: don't move on noisy/duplicate frames; settle after moves; tune gains#452
Slicecam fine-acquire: don't move on noisy/duplicate frames; settle after moves; tune gains#452cfremling wants to merge 1 commit into
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: cd07a1eb4a
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| for ( size_t i = 0; i < nbytes; ++i ) { | ||
| sig = ( sig ^ static_cast<uint64_t>( bytes[i] ) ) * 1099511628211ULL; | ||
| } | ||
| if ( sig == this->fineacquire_state.last_frame_sig ) return; // identical frame; wait for a fresh one |
There was a problem hiding this comment.
When the repeated frame has no detectable centroid, this early return happens before calculate_centroid() and before consecutive_centroid_failures is incremented. In the stale-frame case this code is meant to handle, fine acquisition can now stay running forever on the same bad image instead of climbing the exposure ladder and eventually stopping; this is especially easy to hit with an unrefreshed camera frame or a fixed emulated image. Please either reset/age the signature on acquisition resets or let duplicate frames participate in the no-detection timeout path.
Useful? React with 👍 / 👎.
…fter moves; tune gains Targeted robustness fixes to the slicecam fine-acquisition loop (do_fineacquire), keeping the diff minimal: - Gain 0.7 -> 1.0 (FINE_ACQUIRE_GAIN): a fine-tune step applies the full measured offset; deliberately under-correcting a known offset is never right. - Refuse to move on a noisy solution: only command a correction when the per-axis MAD scatter is within FINE_ACQUIRE_PREC; otherwise discard the batch and re-gather. Previously a move was sent at max_samples regardless of scatter. - FINE_ACQUIRE_PREC default 0.1 -> 0.4 arcsec and made configurable, so good data passes the scatter gate (with 0.1 it rarely did and fell through). - Post-move settle: wait FINE_ACQUIRE_SETTLE_SEC (default 4s) after a TCS move before evaluating the next frame. Time-based, independent of frame cadence. - Reject duplicate frames: a commanded framegrab does not guarantee the image refreshed; processing the same frame twice double-counts samples and can command a second move from a pre-move image (overshoot). Skip a frame whose full-frame byte signature matches the previous one. The signature is reset at acquisition start, and consecutive identical frames are bounded: if the camera is stuck on one frame (or a fixed emulated image) fine-acq stops instead of acquiring forever on a stale image (this early-skip otherwise bypasses the no-detection timeout below). - Centroid Gaussian weighting sigma made configurable (FINE_ACQUIRE_CENTROID_SIGMA), default 1.5 px -- empirically the most robust across SNR and seeing on real SCAM data (was hardcoded 2.0). The optimum scales with PSF width, so it is tunable. Offset measurement/application sign verified unchanged from the proven auto-acq-clean engine (both apply ACAMD_OFFSETGOAL(star - goal)).
cfremling
force-pushed
the
fix/fineacquire-fixes
branch
from
cd07a1e to
821fbb5
Compare
|
Closing as superseded by #455. The minimal config + default tweaks there (FINE_ACQUIRE_GAIN=1.0, prec_arcsec=0.35, goal_arcsec=0.4, and a tunable time-based FINE_ACQUIRE_SETTLE_SEC) proved sufficient on-sky the night of 2026-06-16/17, so the heavier noisy/duplicate-frame + gain-scheduling changes here weren't needed. |
1 participant
Targeted robustness fixes to the slicecam fine-acquisition loop (
do_fineacquire), addressing wrong/overshooting moves seen onmain. Kept intentionally minimal: 5 files, +131/−13.Changes
FINE_ACQUIRE_GAIN0.7 → 1.0. A fine-tune step should command the full measured offset; deliberately under-correcting a known offset is never beneficial here.FINE_ACQUIRE_PREC; otherwise discard the batch and re-gather. Previously a move was sent atmax_samplesregardless of scatter — i.e. acting on an untrustworthy median.FINE_ACQUIRE_PREC0.1 → 0.4″, now configurable. With the old hardcoded 0.1″ the scatter gate was rarely satisfied, so the loop fell through to the move-anyway path.FINE_ACQUIRE_SETTLE_SEC(default 4 s) after a TCS move before evaluating the next frame. Time-based, independent of frame cadence — the move + CCD readout take time, and acting on a mid-move/mid-settle frame yields a wrong correction.FINE_ACQUIRE_CENTROID_SIGMA, default 1.5 px), was hardcoded 2.0. 1.5 was the empirically most-robust value across SNR and seeing on real SCAM data (Commissioning/guiding_20260519, FWHM≈6 px), evaluated against a 2D-Moffat-fit ground truth. The optimum scales with PSF width, so it's tunable.Verified
feature/acq-automation-cleanengine: both measure(star − goal)with cos(dec) and applyACAMD_OFFSETGOAL(star − goal). So this does not change which way the telescope moves — only when and whether a move is issued, plus the centroid weighting.Out of scope (deliberately)
Not compiled locally (no local build available).
🤖 Generated with Claude Code