#18 · stacks/drone

What to build

Plan and simulate a survey polygon with altitude, GSD and image-overlap constraints and produce an auditable georeferenced photo dataset.

Acceptance criteria

Implement boundary clipping, crosshatch/terrain options, camera footprints, trigger timing and flight/endurance estimates. The owned simulator now supports clipped concave polygons/holes, crosshatch and a declared terrain plane; real DEM/collision and measured airframe planning remain open.
Associate exposure timestamps, pose, camera calibration and RTK/PPK/GCP metadata with every image. Synthetic exposure/ACK/pose/calibration records are implemented; correction and control-point status is explicitly unavailable until qualified inputs exist.
Export interoperable photogrammetry inputs and quality reports; distinguish collection from reconstruction. PNG/ODM-style geolocation and collection quality export are implemented; an external reconstruction interoperability campaign and measured survey reference remain open.

Implemented owned simulator slice

Commit 2aad9f3 adds:

Strict versioned survey configuration, concave polygon/hole clipping, crosshatch, calibrated Brown5/crop/nadir footprints, planar terrain following, GSD/overlap/trigger constraints and declared endurance estimates.
Shared calibrated camera geometry with the existing vision localizer, plus exact WGS84 fixed-NED inverse conversion.
Compilation into the owned quad mission with replayable speed, speed-derived task deadlines, raw IMU/GNSS/estimator/control/authority physics, bounded shutter requests and exact scoped ACK/capture-pose records.
Distinct deterministic scene pixels with independently tested landmarks. Plant truth is used for rendering; optical-center geotags use qualified estimator capture poses with explicit initial north assumptions and retained uncertainty.
Camera timeout/scope, raw IMU/position loss, RC takeover, cancellation and incomplete-duration cases. Modeled outputs are neutralized; an airborne release is not an emergency landing.
Private bounded PNG datasets, four-column ODM-style geo.txt, full calibration/pose/correction availability manifest, image hashes and sampled collection reports. Files/directories receive fsync barriers before an exclusive completion manifest. Hashes are unsigned and storage power-loss behavior remains unmeasured.
Native Stacks mapping:plan and mapping:simulate commands, standalone CLI support, and stx /mapping playback that withholds future exposures and final outcomes.

The default nominal survey completes with 32 photos across five lanes; crosshatch collects 63 and declared sloped terrain collects 35. A completed mission can still have a partial image quality report. Reports keep measured GSD, reconstruction error and physical survey accuracy unavailable.

Read mapping contract, ADR 18, and example configuration.

Validation evidence

Local integrated suite before review fixes: 408 Bun tests, 25,604 assertions, zero failures. Final affected suite: 74 tests, 1,851 assertions, zero failures. Thirteen Python bridge tests passed.
Lint and application/runtime/view typechecks passed. Roadmap and parity-integrity checks passed without claiming full upstream parity.
Compiled CLI ran from outside the checkout, including nominal dataset, crosshatch, terrain, camera loss, IMU loss and incomplete-duration exit. Independent Python checks verified PNG CRCs, decompressed scanline sizes and manifest SHA-256 image hashes for all 32 nominal images.
Real browser QA covered nominal/crosshatch/terrain completion, camera-loss partial collection, beginning/end/scrubbing, clear/scenario reset, empty error console and zero horizontal overflow at 320/375 px. Desktop and mobile layouts visually inspected.
Independent Standards and Spec reviews passed after fixing mixed evidence, clock-profile mismatch and speed-derived launch/landing timeouts. Capture clock error propagates through attitude and lever-arm position uncertainty; translation at an upper GSD bound is not claimed as a sharpness certificate.
Exact-head CI passed: 414 Bun tests, 25,631 assertions, 13 Python tests, lint and application/runtime/view types, roadmap/parity integrity, all Linux standalone scenarios including mapping exports/crosshatch/camera loss, and both neutral-only Pico 2 cross-builds.

Remaining work

Keep this issue open for full acceptance:

Measured Pi lens/mount/ROI/timing and physical shutter/ACK integration, including durable request/ACK reconciliation of uncertain captures across restart.
Qualified RTK/PPK/GCP input/provenance and per-image observation import, surveyed absolute heading, common vertical datum and any requested geoid transformation.
Real terrain/DEM and obstacle/collision-aware survey execution, richer rally/transit and measured battery/wind planning beyond the bounded flat-contact quad fixture.
External photogrammetry import/reconstruction campaign, held-out surveyed reference/uncertainty validation, HIL/vehicle survey and hardware storage power-loss evidence.
Rolling shutter/integrated blur and measured image-scale/overlap assessment. Four-corner/grid coverage remains an estimate on a declared plane.

Surface/stockpile processing from known georeferenced inputs is the next software frontier in #19. Collection, reconstruction and volume computation remain separately identified.

Blocked by

#10 Mission engine: run and resume an owned waypoint mission
#8 Hardware adapters: command and verify the first camera or payload
#9 Operations: record, replay and diagnose native missions

Scope and evidence

The implemented evidence is owned simulation and local file/CLI/browser behavior. Physical outputs remain inhibited. It is not firmware SITL, HIL, measured photogrammetric reconstruction or physical survey acceptance. Preserve explicit operator authority and replaceability behind shared runtime/driver contracts.