package/progress.md

DROID-Sim release completion

Scope and locations

• Dataset: ../batch_1st_scene (123 numbered scenes discovered on 2026-11-01).
• Maintained release code and skill: this package directory.
• Per-episode originals and decoded replay inputs: <scene>/trajectory.
• Replay experiments, logs, candidates and temporary outputs: ../replay.
• Source task: /root/.codex/attachments/8bcad202-4603-4baf-897e-45bbcd403941/pasted-text-1.txt.

Requirements / completion evidence

• Manual-review skill written and validated (skills/droid-sim-review/SKILL.md).
• Historical workflow/editor source and point annotations retained with verified hashes in source/; maintained standalone rendering verified for all scenes. The requested skill records safe manual-edit procedures. A new portable GUI is not part of the requested release task; the archived editor is explicitly not certified for cross-frame pose synchronization.
• Dataset inventory and explicit capability assessment: gravity-frame assets and camera, robot/gripper initialization, trajectory replay, common-camera foreground-priority GS rendering; limitations documented in readme.
• Scene-local active rendering/import contracts and portable rendering code: all 123 checked; a minimal four-file isolated scene renders while original dataset/transition reads are denied by a Python audit hook. Recheck after final cleanup.
• GLB-to-USD converter migrated and smoke tested with Isaac Sim 4.5 on scene0001; 3 USD meshes, materials retained, source preserved. Output under ../replay/conversion_smoke/0001.
• Per-scene parameters.yaml: provenance-aware gravity, collision assets/transforms, friction estimates and validation status; 103 unregistered nodes remain explicitly disabled, not represented as solved physics.
• All 123 scenes have matching original DROID metadata and trajectory (533,227,401 original bytes total); SHA-256 checked, original metadata filenames retained. Every released input image matches its source first frame exactly. reports/trajectories.json.
• Replay pilot and documented budget: 657 measured states replayed; bounded scale and 21 rigid-group translation candidates tested in under 10 seconds of automatic placement tuning. Camera/base and accepted scenes unchanged; no candidate passed both contact and visual gates. No batch dynamic grasp validation claimed.
• All 123 input images in media/frame0_inputs.png (3840 x 3120), numbered index alongside.
• Slow all-scene shared-camera orbit GIF using real GS backgrounds: media/frame0_orbits.gif, 123 scenes, 3200x2600, 81 frames, 8.1 seconds; decoded every frame and checked index coverage. File is approximately 324 MiB.
• Dependency-checked cleanup of transition results and scene backups, with required source/annotations and native assets retained; post-cleanup validation follows below. Historical source/vendor/config directories are intentionally not deleted as generated outputs.
• Final per-requirement audit: 123 scene-local contracts, original trajectory hashes, unchanged final GLBs, source archives, complete orbit coverage, converter/replay/skill deliverables and zero remaining scene backups verified. Physical validation limitations remain explicit, not treated as solved.

Current findings

123 numbered scene directories exist. Early and late scene layouts differ. Important gravity transforms/camera overrides originally lived under backup; their active numeric render contracts have now been migrated. A full active-dependency check finds no render/physics/trajectory dependency on backup. Original-input/provenance auditing is complete and all 123 backup directories have been removed. Original trajectory coverage is verified 123/123. No downloads were necessary.

123 portable render/scene.json contracts have been recovered and rendered successfully. Of these, 53 reproduce accepted composites at the near-exact threshold (maximum channel difference <= 1/255, differing-pixel fraction < .0001). Rendering success is not visual or physics acceptance. Reports are ../replay/render_validation*/validation.json. Scene0579 differs on only 0.00001194 of pixels; scene18/19 are pixel-identical. Scene75/224 recovered cameras reproduce accepted foreground layers within the near-exact threshold, but their final composites differ (background and/or compositing policy must be audited).

All missing render contracts are now recovered. Scene6's original final numeric camera was not found: five retained gravity candidates failed image reproduction. fit_archived_camera.py recovered a camera by mesh-depth/SIFT/PnP against the accepted foreground PNG, without moving the GLB. Verification MAE is .002172/255 and alpha IoU .999993; edge differences reach 64/255, so this is explicitly labelled image-estimated, not exact numeric recovery. Scene289's passed GS run was recovered after verifying source image/PLY hashes and the current VGGT camera; its foreground still differs slightly from the archived layer (MAE .386/255). Existing real-GS packages for 0240/0247/0254/0265 were found under batch_1st_44_80_mid_result/diagnostic_failure_review_previews_50_20260730/alignment, checked against current images/native PLYs/VGGT cameras and retained. Their strict-gate failures remain explicit; no automatic acceptance was fabricated. Scene265's original panorama was restored byte-identically from cleanup archive to background. Additional scenes (including 0109) have aligned GS assets but their archived accepted PNG used a 2D background; audit accepted render provenance beyond just presence of aligned PLY.

Several early rerenders differ from accepted composites even though BOTH individual foreground and background layers reproduce. These differences follow the requested foreground-priority policy versus historic cross-layer depth masking. Do not misdiagnose them as camera drift or overwrite accepted images without recording this policy change. The completed provenance audit classifies all 123: 53 near-exact composites; 47 matching individual layers/different compositing; 12 historical background layers pixel-identical to 2D completion; seven minor foreground raster differences; four historical renders referencing different foreground GLB bytes. Report: reports/render_provenance.json.

parameters.yaml generated for all scenes with actual GLB nodes, friction priors, retained collider references, registration evidence, and unknown physical quantities explicit. Initial actual GLB node count is 835 versus 787 historical metadata entries; these are not necessarily one-to-one, and must not be conflated. The exporter understands both obj_N and object_N_geometry_N names. recover_colliders.py indexed 2515 saved sim-ready input meshes and recovered 210 matching collision sets into scene-local physics/colliders, preserving source hashes and a verified input-to-final affine transform (maximum vertex error <=2e-5 scene units). Registered collider coverage is now 732/835 final mesh nodes; 103 remain unmatched/missing. Full records: reports/collider_recovery.json. No dynamic physics is claimed validated, and missing/mismatched colliders are disabled. The 123-scene parameter export completed again without exceptions after this recovery; see reports/physics_parameters.json.

Isaac scene0001 observed-state replay now completed all 657 samples using the official nominal FR3+hand in 25.315 seconds, maximum immediate joint assignment error 1.63e-9. ../replay/pilot/0001 contains frame0/last-frame PNGs, an initial USDA, replay_trace.npz and replay_status.json. The first pilot's obstructing robot was caused by stale render/default import pose and an unmoved fixed-joint anchor: placing the USD assembly before reset and stepping physics before capture fixed it. Actual base position matches the calibrated transform within ~1e-8 meters, and the table center is unobstructed. This is still NOT a validated metric robot-scene alignment or grasp test: scale is an explicit 1 m/unit hypothesis. The official Franka description, tag 0.5.1 / commit a72e2061021648371d0c9e93de11bf33ca83b6f7, is packaged at assets/franka_fr3 (22 files, approximately 27 MB), with relative meshes, Apache-2.0 LICENSE/NOTICE and hashes. The source checkout is under replay; no remote repository was modified. The model is nominal, not serial-calibrated, and cannot be called the exact model for the dataset's Panda/ambiguous serials. Isaac shutdown can return exit zero after errors: inspect replay_status.json, not just process status.

audit_robot_kinematics.py verified all 123 trajectories against nominal flange FK. Held-out odd-sample residuals after an even-sample constant tool-frame fit are at most 4.07e-8 m and 5.14e-6 degrees. The fitted offset is effectively identity: the recorded Cartesian frame is the flange, not the fingertip. This checks conventions/data consistency, NOT independent physical calibration. Report: reports/robot_kinematics.json.

The bounded scene0001 placement trial is recorded in ../replay/placement/0001/placement.json. With explicit +Z fingertip-offset hypothesis 0.1034 m, first gripper closure at sample110 / 7.745s and the towel as target, the best projection-preserving scale hypothesis 0.84358 m/unit still leaves 0.14007 m contact residual and 92.49-pixel discrepancy. It was rejected against a 0.03 m residual gate; accepted geometry/camera/base were unchanged. The numerical fit took 0.029 seconds; no expensive iterative placement was launched. An unconstrained nearest-object search incorrectly preferred the container, so it must not be treated as grasp alignment. Further automatic contact optimization is still unvalidated and must stay within the five-minute average budget.

Rendering policy is foreground-priority conditioned rendering: depth resolves visibility within mesh foreground and within GS background, then foreground alpha takes priority over GS. It is not a claim of physically validated cross-layer metric depth ordering. Automated alignment/physics failures remain recorded even when human visual review passed.

Next actions

Requested packaging, cleanup, media and replay-placement pilot are complete. Further metric calibration, missing collider reconstruction, dynamic contact/grasp validation or a new portable GUI would be follow-up work, not certified by this release. No automatic batch placement is applied from the rejected pilot.

Latest checkpoint

All seven validation batches finished: 106 + 7 + 2 + 2 + 1 + 1 + 4 = 123 scenes. The optimized persistent orbit renderer was compared against independent camera-gauge rendering at 5 degrees: MAE .0000926/255, maximum channel difference 1/255. The completed full batch has 81 frames at 10 fps, triangle motion 0 -> +10 -> -10 -> 0 degrees around each scene's foreground bounds center and gravity Z axis. All layers use the same camera and real GS. ../replay/orbit_renders/status.json confirms 123/123; media/frame0_orbits.json indexes all tiles. Refreshed GIF decoded successfully (81 frames, 8100 ms, 339758565 bytes). This is visual novel-view preview, not a guarantee of complete geometry at unseen viewpoints.

Final completion audit — 2026-11-01

Post-cleanup validation passed 123/123 (reports/release_after_cleanup.json). A separate final read-only audit re-hashed all accepted GLBs against earlier validation, all original trajectory/metadata files against their source-copy hashes, and retained source/annotation archives. All matched. All 123 scene backup folders are absent; every parameters/render contract remains present. All package Python sources parse. Input PNG verifies, GIF has 81 frames, and orbit scene IDs exactly cover the 123 numbered release scenes.

Rigid-group placement pilot: 21 candidate fractions from 0 to 1 in 9.488 seconds. At 1 m/unit, nearest towel contact residual starts at .200078 m. The full shift makes the geometric residual zero but target IoU falls to .181906. Even the first sampled 5% shift fails the .95 target-IoU gate (.923418). Only the zero shift passes both visual gates, so accepted assets, camera and robot base remain unchanged. This is a useful negative result under the requested time cap, not a claim of successful contact. Files: ../replay/placement/0001_translation. The earlier scale-only trial took .029 seconds. The measured-state Isaac replay is separately verified for all 657 samples; dynamic grasp remains untested.

Native Marble PLY/panorama in accepted scene folders and original raw episodes are unchanged. Verified duplicate copies inside retired runs have been removed. No accepted GLB or root layered-composite PNG was overwritten. No external GitHub backup changes are authorized or planned.

Cleanup and dependency checkpoint

validate_release_contracts.py --mesh-check passed 123/123 scenes. It verified active scene-relative paths (no backup dependency), Gaussian PLY headers, shared-camera matrix identities, original trajectory hashes, final-node transforms and 47,296 collision-part files for mesh readability. There remain 103 unregistered final nodes. For scene0013, for example, saved collision-input meshes have different vertex counts and geometry from final assets; treating these as corresponding vertices would be incorrect. None of these checks certify Isaac collision cooking, bowl cavities, mass, metric scale or grasp success.

retain_workflow_inputs.py saved 203 workflow source files (727,672-byte compressed archive) and 199 point-annotation/inventory/mask inputs (149,246-byte compressed archive), then read back and verified every SHA-256. This includes five supplemental support-bridge masks. Archives are under source, with full retention report reports/retained_workflow_inputs.json; credentials/environment files and generated previews are excluded. Archived source is historical and not silently advertised as portable runtime code.

First cleanup: 30 obsolete preview/log directories were removed from transition, totaling 3,876,998,799 logical bytes. Any Gaussian PLY removed with a preview was first hash-verified against the retained release copy. Five candidate folders with other protected inputs were initially skipped. See reports/cleanup_previews.json.

Main transition cleanup then removed 10 generated-result/tmp/review/input-copy folders totaling 67,432,340,385 logical bytes (reports/cleanup_transition.json). Fourteen native PLY/panorama duplicates were hash-verified against retained scene files. Nine misleadingly named background_marble.ply files were verified from their generation manifests and hashes as transformed gravity-frame derivatives, not native Marble originals. Source code, vendor libraries and configurations remain in the old workspace; required source is also archived in package. Nested Python bytecode caches were removed separately; no primary source files were edited.

All 123 scene backup directories were removed after the main transition cleanup and dependency/provenance checks: 1,116,526,389 logical bytes (reports/cleanup_backups.json). Thirty-six input/runtime-source files from backups were preserved in the verified 17,884-byte source/backup_inputs.tar.gz. A late-discovered legacy runtime directory was initially blocked from deletion by safety review. A full read-only inventory established that it contained 78 generated JSON/PNG/NPZ/log/lock files, no code, no symlinks and no native assets. Four input JSONs were preserved, the cleanup gained a strict file-type guard, and the re-reviewed operation removed these 26,332,561 bytes (reports/cleanup_transition_runtime.json). This small residual cleanup occurred after the backup phase; no backup-only dependency was removed without migration.

Scene0003 camera recovery: current and archived GLBs were byte-identical, but the earlier recovered camera yielded foreground MAE 6.538392/255. Depth-backed SIFT/PnP recovered an image-estimated camera with MAE .003399/255 and alpha IoU .999962, passing the explicit estimated-camera gate. Only the newly introduced render contract changed, not accepted geometry or PNGs. Evidence is under ../replay/camera_recovery/0003_fit and 0003_verification; this is not exact numeric-camera recovery. Scene0063 fitting was rejected for insufficient inliers. Scenes0126/0204 retained-camera candidates also failed reproduction and were not published. Their historical manifests, and scene0063's, reference different GLB hashes from current final assets, so forcing current cameras to those old images is unjustified.

All-scene GIF was refreshed for scene0003, then all 81 frames decoded again: 3200x2600, 8100ms, 339,758,565 bytes. Every orbit cache camera-contract hash matches the current 123 scene contracts; no stale-camera tile remains.

Post-cleanup active-contract verification passed 123/123 again (reports/release_after_preview_cleanup.json). A separate runtime test copied/hardlinked only four inputs into an isolated temporary scene: render contract, GLB, aligned GS PLY and input image. It denied Python file reads into the original dataset and transition workspace before importing the renderer, then produced a real-GS composite at 320x180 in 8.854 seconds without cache. Evidence: ../replay/portability/0001/portability.json; temporary scene removed after the test. This supplements the all-scene path audit; it is not a clean-machine installation test or an OS-level syscall sandbox.