The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Error code: FileFormatMismatchBetweenSplitsError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Complete original manuscripts with measured 3D additions
All 19 original main-paper figures and 23 supplement figures retain their colours, arrays and narrative order, with ten recovered original gallery sheets. Main Figure 1 adds matched measured 3D terrain before/after its translucent target and the same-state online proposal. Original source PDFs remain byte-identical. The complete original arguments remain with an explicit correction ledger; unsupported exact, flight-safety, variable-speed energy and causal-control interpretations are qualified rather than presented as new validated results. The named paper has 35 pages; the expanded appendix has 460 pages.
- Reproduced original main paper
- Reproduced original supplement
- New figure/argument/state appendix
- Expanded visual study
- Complete validation record
- Complete expanded appendix
- Combined author archive
- Anonymous TMLR working manuscript
- Saved-state ZIP
- Models, source, native arrays, scenes, films and reviews
Three new DA2 Small adapters use nine evolving powered states and nineteen rasters. Each starts from pinned depth foundation weights with fresh seed 0/1/2, optimizer and task initialization, completes 12 epochs/768 updates, and saves selected plus final optimizer/scheduler/RNG payloads. Fresh 96-terrain support yields 384 queries with disjoint 64/16/16-terrain splits. All 192 reserved and 288 fixed measured native outcomes are saved. Mean reserved/measured MAE: 0.127941/0.209337. The measured result is worse than previous narrow DA2 0.162815 by 0.046522; these fits are diagnostic, not promoted. Earlier DA3 0.142524 remains stronger. Training distributions differ and the measured panel was already inspected; this is not a causal state-support ablation or new product-held-out geography.
Independent review reproduces all 480 native outcomes, checks the algebra of 1,054,122 winning witnesses, and densely samples 46 complete paths. The finite constructive reference is not an optimized global nine-dimensional HJ solution or contact/controller test. No 10–12 evolving-state teacher is claimed. The eight-raster gliding interface adds available altitude from observed geometry and supplied telemetry; its saved initialized adapter is explicitly untrained and ignores its zero-kernel eighth channel initially. Original trained gliding results remain seven-input. Heading and speed are supplied aircraft state, not guessed from terrain.
All original PDFs, all prior scientific studies, the changing-state movies, Viser views, real-video depth diagnostics and enlarged 300-world gallery remain. The immediate parent 28 GB release and earlier 121 GB campaign are pinned in STATE.json. This additive capsule preserves all new state without duplicating every prior checkpoint. Every local ZIP member is decompressed and SHA256 checked; remote object/PDF verification follows separately. No external website deployment, journal submission or flight-safety certificate is implied.
Corrected visual story3 — original RGB narrative and fresh seed repeats
The preferred actual draft8/supplement palette is restored with66% transparent red/yellow/green terrain fields, blue time/budget and purple errors. Physically dimensioned white triangles, dotted history, edge legends/metre bars and larger panels make the numerical meaning visible. Every original PDF/reference and the rejected prior iteration remain preserved. The named paper has26pages; the complete research appendix has369pages.
- Named author paper
- Complete research appendix
- Combined author archive
- Anonymous TMLR working manuscript and focused appendix
- Complete saved-state ZIP
- Models, arrays, source, scene files, films and reviews
All nine fresh adapters start from pinned depth foundation weights: DA2 full, DA2 angles hidden and DA3 full, seeds0/1/2, with paired epoch orders,19channels, 224square input,12epochs/768updates and float64 AUROC validation selection. Reserved/measured mean MAE across all3seeds: DA2 0.12165/0.16282; DA3 0.10528/0.14252; hidden-angle DA2 0.17845/0.26842. DA3-minus-DA2 three-seed intervals include zero. These repeat optimization conditional on already inspected fixed panels; they add no product-held-out geography. Historical single-seed results remain separate. All selected/final restart payloads,108epochs,1440native query outcomes and exclusions are saved.
Thirty-two actual1Hz current-state queries change the Moon/Mars film fields. Dotted trails include only past states. Every query tests new8–54s future transfers, adding the positive remaining original duration; the zero-velocity 2m endpoint is explicitly labelled as a new hypothetical future-transfer query. It is not touchdown or a shrinking residual-episode envelope. Severe late raw student failures remain visible and the web slider/error map is synchronized. Independent checks cover800,936witness cells and72dense paths. Original primary campaign results, failed developmental neural teachers,103sample real-video relative-depth diagnostics and all300native gallery maps remain preserved.
All local ZIP members are decompressed and SHA-256 checked; remote object and full-PDF verification follows separately. The large named archive is separate from an anonymous journal supplement; no website deployment or OpenReview submission was performed. Prior dataset objects are unchanged.
Visual story 3 — heading and nine-state approaches
The next iteration preserves all original draft/supplement pages and adds muted direct-terrain colours, explicit heading, aligned HJ/analytic/ student/error maps, a trained 19-channel DA3/DA2 powered pilot, and enlarged galleries. The public TMLR-format draft has 23 pages; its appendix has 333.
- Paper
- Detailed appendix
- Complete new-stateZIP
- Source, models, raw arrays, Viser, renders and media
- Source branch
The one-seed nine-state constructive pilot tests a two-second ballistic actuator alignment followed by checked powered cubic approaches to a 2 m airborne gate. DA3 Large has 354.81M retained/88.27M trainable parameters: reserved MAE 0.09914, measured MAE 0.14180, isolated A6000 forward 30.66 ms. DA2 is 25.99M with MAE 0.12842/0.17522 and 9.00 ms; hiding angles yields 0.17674/0.26338. Measured transfer uses 96 fixed queries on 12 reused windows/two products, only two joint angle states. This is not independent product-held-out generalization, a DINOv3 experiment, global HJ optimality or touchdown safety.
Full Moon18 s and Mars12 s checked paths plus2 s holds use fixed initial targets. Eight later-state diagnostics retain severe late model failures, including32016 reference-negative calls among32017 late Mars calls and an unwitnessed selected proposal. These are illustrations, not new primary aggregates or a flight controller. The same source terrain, numeric drapes and query poses drive high-resolution Blender films and standalone Viser scenes. The entire 205 s NASA montage also has 103 pinned pretrained DA3 relative-depth samples; this perception-only diagnostic does not supply a metric DEM or landing evaluation. Every local ZIP member is decompressed and SHA-256 checked. Independent numerical/model/source/layout reviews and browser checks accompany the files. Pinned remote-object/download verification follows as a separate proof. Earlier full campaign releases and original paper files remain unchanged.
Visual story 2 — heading and nine-state approaches
The next iteration preserves the preferred original figure narrative and adds a rebuilt measured 2D/3D opener, explicit heading, aligned HJ/analytic/ student/error maps, a trained 19-channel DA3/DA2 powered pilot, and enlarged galleries. The public TMLR-format draft has 18 pages; its appendix has 249.
- Paper
- Detailed appendix
- Complete new-stateZIP
- Source, models, raw arrays, Viser, renders and media
- Source branch
The one-seed nine-state constructive pilot tests a two-second ballistic actuator alignment followed by checked powered cubic approaches to a 2 m airborne gate. DA3 Large has 354.81M retained/88.27M trainable parameters: reserved MAE 0.09914, measured MAE 0.14180, isolated A6000 forward 30.66 ms. DA2 is 25.99M with MAE 0.12842/0.17522 and 9.00 ms; hiding angles yields 0.17674/0.26338. Measured transfer uses 96 fixed queries on 12 reused windows/two products, only two joint angle states. This is not independent product-held-out generalization, a DINOv3 experiment, global HJ optimality or touchdown safety.
The same source terrain, numeric drapes and query poses drive high-resolution Blender films and standalone Viser scenes. The entire 205 s NASA montage also has 103 pinned pretrained DA3 relative-depth samples; this perception-only diagnostic does not supply a metric DEM or landing evaluation. Every local ZIP member is decompressed and SHA-256 checked. Independent numerical/model/source/layout reviews and browser checks accompany the files. Pinned remote-object/download verification follows as a separate proof. Earlier full campaign releases and original paper files remain unchanged.
Current visual narrative — 2026-10-06
The next iteration is available: a 9-page TMLR preprint with a large opening 2D/3D hero and a 213-page detailed appendix preserving the entire prior 206-page scientific record.
- New paper
- New detailed appendix
- Complete iteration ZIP
- Iteration files, models, website and media
- Source branch
Measured Apollo 11, Jezero and contemporary Hudson scenes use explicit source/geometry contracts. Three new 1080p camera films and standalone Viser scenes accompany the paper. The website retains the long measured terrain film and official NASA yt-dlp footage/relative-depth diagnostic. Stage and heading panels share one saved numerical HJ computation.
The displayed full-state 364 px powered model is selected by validation among three matched fits: 25.84M parameters, 11.64M trainable in this setup, 98.7 MiB FP32 checkpoint. Best and resumable last states are included. The separate 7D/9D development teachers failed validation; no trusted real-terrain labels are generated with them. All primary results, analytic baselines, prior release pins and scientific PDFs remain preserved. Context glyphs and hypothetical queries do not reconstruct Apollo or Ingenuity flights or establish physical landing safety.
The full iteration ZIP and member manifest were independently checked locally. Bounded independent paper/source/media/browser reviews are included. The preceding campaign and its pinned artifacts follow below.
ReachDown research campaign artifacts
The eight-hour campaign is complete. The linked final source snapshot contains the final paper and review records.
Read the paper and detailed appendix. These PDF bytes are extracted from the immutable final-research-source-recovery-bound-20261006.zip; their checksums and publication commits are recorded in artifact-index.json. The source ZIP retains TeX, figures, experiment protocols, review reports, website source, and media.
This publication preserves 63 independently verified ZIPs,
totaling 121,149,684,851 bytes. The inventory below is pinned to
Hugging Face commit 2596b3b337ea3d49dedd7cb26a13bc3935d48dd7. Every listed ZIP has a matching
.manifest.json and .sha256 companion at that commit and an internal STATE.json.
Remote Git-LFS hashes, byte sizes, and independently downloaded companion bytes match
the closed local archives. Every listed ZIP also has an independent complete-object SHA256,
every-member SHA256/CRC, exact member-set/size and internal-state proof, bound in
complete-integrity.json.
The catalog composes those completed reads; it does not reread all payloads when generated.
These integrity checks do not by themselves validate a research claim. Scientific review
reports are retained in the source and named evidence archives.
Research scope
The powered extension uses a seven-coordinate vacuum vehicle state and a constructive trajectory family ending at an airborne gate. Its 17 input channels include vehicle/state conditioning. The primary comparison retains all six 224-pixel fits and compares presence of vertical velocity; both arms retain mass and start from the same pretrained backbone. The analytic surrogate has lower native map MAE than every full-state primary fit. Measured terrain results are heterogeneous by body and optimizer seed. A positive reference is a reduced-model witness; a negative reference is unresolved physical feasibility. The paper reports proposal coverage and utility as well as conditional quality. These artifacts do not establish touchdown, obstacle completeness, flight safety, probability calibration, or a benefit from the original foundation-model pretraining.
The expanded campaign retains all six 364-pixel fits and both seed-zero weight-one controls, alongside the six primary fits. Every fit completes 24 epochs and 9,816 updates. Larger inputs increase computation; the comparison does not match FLOPs or wall time. Objective sensitivity is a separate one-seed study. All original native predictions, per-body outcomes, exact query AUROCs, proposal coverage and observed checks remain available. Primary synthetic improvement does not imply uniform measured-terrain or policy improvement.
All 5,292 original finite-family references replay bitwise before 192/384-interval refinement. Fixed model scores and actions are compared at each reference level, with level-specific and common-defined query denominators. This finite census is not a convergence, optimal-reachability or physical-safety theorem. Complete offline final-model/AdamW/scheduler/CPU-RNG reconstruction passes for all 14 fits; CUDA RNG payloads are preserved without an on-device restoration claim. The separate same-backend checkpoint-to-native-prediction replay covers 16,072 maps with maximum probability difference 0, 0 cell-call changes and 0 emitted-action changes. Bitwise, tolerance and action-parity results are reported separately. These restoration/integrity checks do not establish interrupted-training, cross-device or physical accuracy equivalence.
The separate 28-forward CPU sample uses only two underlying lunar synthetic scenes per fit. Maximum probability drift is 0.00188291 and two native cell classifications change, while the sampled argmaxes/proposals/emitted actions stay unchanged. It does not establish cross-backend cell parity. Original learned timing fixtures all abstain (1,600 measured requests). Supplemental first-lexicographic called proposals have separately measured CPU verification times on six fixed cases; all six accept. Neither timing protocol establishes population, rejected-case, onboard or end-to-end latency.
The final frozen research inventory is covered by pinned published ZIP members and freshly reread file hashes. Isolated final-source recovery reproduces full paper text/page counts and selected-page pixels, followed by one explicitly limited offline CPU fixture. The completed source commit is independently recovered from initial plus incremental Git bundles. The exact records and their scopes are bound in the artifact index's completion evidence. Publication and closure attestations after the inventory cutoff are separately retained. The completion proof record binds their exact paths and hashes; these proof files are also published directly.
Measured lunar and martian terrain, orbital texture, reconstructed Mars 2020 kinematics, official camera footage, rendered geometry, and trained-model replay have separate provenance. The three-minute textured movie uses measured 2 m terrain heights with same-pair orbital grayscale imagery. Fine visible texture is not extra recovered terrain geometry. The camera-video depth visualization is relative depth, with no metric scale or temporal reconstruction claim. Full videos, original sources, frame records, meshes, and manifests are retained in the corresponding archives.
Download and restore
Download a ZIP together with its two companions. Check the ZIP against the .sha256
file, then inspect STATE.json or the external manifest before choosing extraction paths.
Archive prefixes are recorded explicitly and differ between source, data, checkpoint,
and media archives; avoid flattening them or overwriting a newer source snapshot with an older one.
The source ZIP restores a reach-down/ tree. The original dataset/checkpoint ZIPs also
use that prefix. New campaign archives preserve their named data/run directories.
After restoring the desired data, use docs/data/research-symlinks-20261005.json and
scripts/restore_research_links.py to restore the recorded historical aliases.
initial-git-recovery.zip retains the original tracked source, binary working-tree patch,
and parent Git bundle. See docs/release-20261005.md for capture limitations and submodule provenance.
Each completed-fit ZIP contains best.pt for inference and last.pt with the final
optimizer, scheduler, and RNG state, plus split, history, environment, source snapshot,
arguments and job information. Checkpoint state preservation does not imply a tested
bit-exact interrupted-training resume with persistent data-loader workers.
The pinned original depth backbone is preserved separately.
To preview the restored website, serve the repository root and visit /website/;
media links use the neighboring /media/ directory. See website/README.md.
Verified immutable archives
All links below select the recorded commit. Sizes are compressed archive bytes.
Licensing and source
Per-artifact source, model, and data licenses remain attached to those assets. The repository
license does not relicense third-party data or weights. The Nerfies website adaptation is
CC BY-SA 4.0; its upstream commit and license are retained in website/.
NASA/LROC/HiRISE source URLs and product metadata are in the data manifests.
Source branch: research/heading-planetary-paper-20261005. GitHub Releases are deferred; these Hugging Face ZIPs are the requested artifact destination.
- Downloads last month
- 9,383