DEEPSIFT

Autonomous downlink research for bandwidth-constrained missions.

Not every bit deserves the trip to Earth.

On a held-out Curiosity Navcam interval, position-based traverse sampling with a stereo-safe progressive scheduler kept 1/4 of traverse frames at full quality — 26.1% of full-quality traverse bytes — while every archived traverse frame stayed within 5 m of a kept frame (at most 2.05 m) and no stereo pair was broken.

HELD-OUT TESTCuriosity Navcam · sols 950–979· Scheduler V3 + POSITION at 1/4 retention · 12 traverses · 679 framesPRE-REGISTERED CRITERION · PASS

26.1%
of full-quality traverse bytes: 100%
5 m spatial coverage: 2.05 m
maximum distance to a retained frame: 0
broken stereo pairs

source · artifacts/phase3_final/20260926T174018-phase3-final-test-e4e2/results.jsonconfig 6f35d7fc205c… committed (96e0694) before the first test image was downloaded

▶ Replay the held-out test Research & paper What didn’t work

01The problem

A rover collects more imagery than it can send home.

Every image a Mars rover takes has to wait for a relay pass to reach Earth, and relay capacity is finite. When storage and bandwidth run short, something must decide what goes first, what is sent as a thumbnail and what waits. That decision is usually made by people on the ground, a sol at a time.

DEEPSIFT asks a narrower, measurable question: which signals actually help decide what should be transmitted — and which only look helpful on the data they were tuned on.

Test interval: 25 active sols
Navcam acquisitions: 1,194
Archived products: 4,342
Full-quality downlink: 311 MB

02What DEEPSIFT does

A small, inspectable pipeline.

The released path uses rover geometry and a byte-accurate scheduler. Nothing in it needs a network, a cloud model or a learned score.

01
Archived observations
Curiosity Navcam raw EDR from the PDS Imaging Node, with URLs, labels and SHA-256.
→
02
Sequence analysis
Acquisitions grouped by capture clock; stereo pairs kept together; traverses located with PLACES.
→
03
Position-aware selection
Farthest-point sampling on rover position keeps frames that cover the traverse.
→
04
Stereo-safe scheduler
Scheduler V3 fills thumbnail pairs, then compressed pairs, then full pairs — never one eye alone.
→
05
Bandwidth-constrained downlink
Coverage can only grow as the budget grows (monotonic by construction and by test).

03What we tested

Most of what we tried did not earn a place.

Each branch was tested against pre-declared rules on data it was not tuned on. The negative results are part of the result.

What didn’t work

Jev semantic rankingDROPPhase 2
AUROC difference vs deterministic rules +0.0056 (95 % CI -0.020 to +0.037); 1,500 live calls; pre-registered stop rule → discontinued for ranking
source · apps/web/data/home-summary.json · run 20260925T094903-jev-v3-pilot-84af
Telemetry image rankingDROPPhase 3.1–3.2
image-independent by construction (0 response to controlled visual change); at 1 % under Scheduler V3 it made 7 acquisitions usable at 1 rover position — the fewest of six strategies
source · artifacts/phase3_2/20260925T135920-phase3.2-dev-corrective-ceb8/results.json
pHash representative selectionDROPPhase 3.3
constrained pHash false-merge rate 0.15–0.19 on validation1 (limit 0.05); ≤ 1.03× traverse compression
source · artifacts/phase3_3/20260926T142033-phase3.3-val-2a31/results.json
QUALITY_V2 as a priority modifierDROPPhase 3.3
false-positive rate 3.6% in development → 10.6% on validation1; kept only as a diagnostic QUALITY_SUSPECT flag
source · artifacts/phase3_3/20260926T142033-phase3.3-val-2a31/results.json
Embedding-assisted selectionNO MEASURABLE ADDED VALUEPhase 3.4 + final test
visual-change gain over POSITION at 1/4: development +0.044, validation1 +0.004, validation2 +0.001, test +0.005 (pre-registered meaningful threshold 0.020)
source · artifacts/phase3_final/20260926T174018-phase3-final-test-e4e2/results.json

What generalized

Rover-position sampling (POSITION)KEEPall four periods
5 m coverage 1.000, 1.000, 1.000, 1.000 across development, validation1, validation2, test
source · artifacts/phase3_final/20260926T174018-phase3-final-test-e4e2/results.json
Scheduler V3 (stereo-safe progressive)KEEPall four periods
0 monotonicity violations and 0 single-eye stereo pairs in development, validation1 (0), validation2 (0) and test (0)
source · artifacts/phase3_final/20260926T174018-phase3-final-test-e4e2/results.json

The simplest signal that generalized was rover position. More complex signals — a semantic model, telemetry context, perceptual hashing, a learned quality detector and image embeddings — either failed validation or added nothing measurable outside development.

04Generalization

One result, four separate periods.

Development, two validation periods and a final held-out test, each reported on its own. The test interval was frozen before any image was downloaded and used once.

Bytes retainedfinal-test criterion ≤ 0.35

Developmentsols 412–430

0.266

Validation 1sols 779–820

0.265

Validation 2sols 1100–1129

0.270

Held-out testsols 950–979

0.261

5 m spatial coveragefinal-test criterion ≥ 0.90

Developmentsols 412–430

1.000

Validation 1sols 779–820

1.000

Validation 2sols 1100–1129

1.000

Held-out testsols 950–979

1.000

Largest distance to a kept framefinal-test criterion ≤ 10 m

Developmentsols 412–430

2.64 m

Validation 1sols 779–820

2.78 m

Validation 2sols 1100–1129

1.94 m

Held-out testsols 950–979

2.05 m

Broken stereo pairsfinal-test criterion = 0

Developmentsols 412–430

Validation 1sols 779–820

Validation 2sols 1100–1129

Held-out testsols 950–979

Scheduler V3 + POSITION at 1/4 of traverse frames. Periods are reported separately, never pooled. The vertical tick is the final-test limit; it was pre-registered for the test only. Development, validation 1 and validation 2 were recomputed by the frozen final-test runner. Source · artifacts/phase3_final/20260926T174018-phase3-final-test-e4e2/results.json → four_period

Visual-change gain of POSITION + EMBEDDING over POSITION (1/4 retention)95 % sequence bootstrap

Development

+0.044

Validation 1

+0.004

Validation 2

+0.001

Held-out test

+0.005

Dashed line: the pre-registered “meaningful” threshold (+0.020). The development gain did not persist on either validation period or on the held-out test, so embeddings get no credit in the primary claim. Visual-change coverage is a MobileNetV2 proxy, not a scientific judgement.

05How we got here

Not one benchmark run.

Every stage is a git tag. Configurations and pass/fail rules were committed before the data that tested them.

Phase 1 · telemetry triage prototype
v0.1-baseline · 90c7f4a · 2026-09-24 15:23
Phase 2 · Jev semantic decision evaluation
phase2-complete · 58725e0 · 2026-09-25 06:19
Phase 3 · Navcam downlink baseline
phase3-baseline · 577d5be · 2026-09-25 09:25
Phase 3.1 · controlled visual tests
phase3.1-complete · af661c6 · 2026-09-25 09:41
Phase 3.2 · stereo-safe scheduler
phase3.2-complete · 6b913fa · 2026-09-25 10:07
Phase 3.3 · config + rules frozen before validation1 download
phase3.3-config-frozen · 6bce97d · 2026-09-26 09:44
Phase 3.3 · validation1 (779–820)
phase3.3-complete · e78aaff · 2026-09-26 10:24
Phase 3.4 · claim + decision rule pre-registered
phase3.4-config-frozen · 566c323 · 2026-09-26 12:31
Phase 3.4 · validation2 interval frozen from listings
phase3.4-validation2-frozen · f800550 · 2026-09-26 12:31
Phase 3.4 · fresh validation (1100–1129)
phase3.4-complete · 33ef7d8 · 2026-09-26 12:54
Final test config frozen before download
phase3-final-test-config-frozen · 96e0694 · 2026-09-26 13:03
Final held-out test (950–979)
phase3-final-test-complete · fdc2122 · 2026-09-26 13:41

06Mission replay

Watch the held-out traverses replay.

HISTORICAL REPLAY · sols 950–979

Final test replay →

Rover path, every archived traverse frame, the frames POSITION keeps, the bandwidth meter, 5 m coverage and stereo state — SEND ALL vs POSITION.

REMS / RAD · STATIC ARCHIVE

Phase 1–2 telemetry study (archive) →

The earlier telemetry-triage study. Its interactive mission-control replay needs the local pipeline API (npm run demo) and is not part of the public site.

07Research

Read the method, the results and the limits.

PAPER · REPORTS · FIGURES

Research overview →

Technical paper, phase reports, methodology, architecture and figures.

HASHES · TAGS · SEEDS

Reproducibility →

Periods, config hashes, commits, seeds, dataset hashes and the run order.

READ FIRST

Limitations →

Survivorship bias, simulated compressed tier, PLACES interpolation, one rover, one camera.

08Reproducibility

Frozen before it was tested.

Final test ordering (from git and the file system)

config committed2026-09-26T13:03:26-04:00

first test image written2026-09-26T17:03:50

config before downloadyes

final config hash6f35d7fc205cdffe4a97…

Artifact structure

config/phase3_final_test_config.json   frozen criterion + source hashes
data/splits/                            dev / validation / test intervals
data/manifests/navcam_*.json            URLs, labels, SHA-256 per product
artifacts/phase3_final/<run>/           results.json · run_manifest.json
docs/release/science-artifacts.json     70 artifact hashes