diff --git a/.gitattributes b/.gitattributes index 07905d7284891a6c080911f2422107c92d982cdb..b882705d8d68c7ff8e90b2476e9bc243c903c128 100644 --- a/.gitattributes +++ b/.gitattributes @@ -759,3 +759,7 @@ sessions/session_20260921_225744_574043/preprocessing/c6a-input-preparation-v1/n sessions/session_20260921_225744_574043/preprocessing/c6a-input-preparation-v1/native/appearance_train_v1/model/point_cloud/iteration_10000/point_cloud.ply filter=lfs diff=lfs merge=lfs -text sessions/session_20260921_225744_574043/preprocessing/c6a-input-preparation-v1/native/appearance_train_v1/model/point_cloud/iteration_7000/point_cloud.ply filter=lfs diff=lfs merge=lfs -text sessions/session_20260921_225744_574043/preprocessing/orange-belt-prior-v1/native/rerun/appearance_spline_prior.rrd filter=lfs diff=lfs merge=lfs -text +sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_train_v1/input_case/session_20260921_225908_522335/observation.ply filter=lfs diff=lfs merge=lfs -text +sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_train_v1/input_case/session_20260921_225908_522335/points3D.ply filter=lfs diff=lfs merge=lfs -text +sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_train_v1/model/point_cloud/iteration_10000/point_cloud.ply filter=lfs diff=lfs merge=lfs -text +sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_train_v1/model/point_cloud/iteration_7000/point_cloud.ply filter=lfs diff=lfs merge=lfs -text diff --git a/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/APPEARANCE_REVIEW_c6a_v1.md b/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/APPEARANCE_REVIEW_c6a_v1.md new file mode 100644 index 0000000000000000000000000000000000000000..ab3eb2e92c93dec9bee5eae373c729e365749879 --- /dev/null +++ b/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/APPEARANCE_REVIEW_c6a_v1.md @@ -0,0 +1,31 @@ +# 225908: static appearance review + +**Status: reference training and agent review complete; owner appearance accepted with the disclosed limitations on 2026-10-29.** + +This appearance uses your accepted 328–560 masks and accepted initial orange prior. It is trained only on the four initial images (converted328/native329), with the unchanged reference recipe: 10,000 loop iterations and 9,999 Gaussian optimizer updates. No physical fitting has run. + +## Review views + +[Rotation video](appearance_train_v1/review/novel_view.mp4) · [eight sampled rotation views](appearance_train_v1/review/inspected_orbit_samples.png) · [loss curves](appearance_train_v1/review/loss_curves.png) + +- [Camera 12540311: observed/rendered RGB, mask/alpha and depth](appearance_train_v1/review/iteration_10000/view_0_sheet.png) +- [Camera 12933256: observed/rendered RGB, mask/alpha and depth](appearance_train_v1/review/iteration_10000/view_1_sheet.png) +- [Camera 17736741: observed/rendered RGB, mask/alpha and depth](appearance_train_v1/review/iteration_10000/view_2_sheet.png) +- [Camera 23007103: observed/rendered RGB, mask/alpha and depth](appearance_train_v1/review/iteration_10000/view_3_sheet.png) + +The visible orange ring and tape align across the four views. **The hidden grasp region remains incomplete and fuzzy**, with a gap and diffuse fringe also visible during rotation. This is an observed limitation of the trained appearance; it has not been filled or rerun with changed settings. The owner accepted this limitation on 2026-10-29. The raw residual sheet compares against the full scene and is not the masked training objective. + +## Saved evidence + +- Initial hybrid seeds: 67,984. Final saved Gaussians: 73,252; reference export after strict opacity >0.1: 64,142. +- One isotropic scale column; final spherical-harmonic degree3. +- Mean same-view source total: 0.005388009 initially → 0.000197463 at iteration10000. This averages the source full-image losses on the four training views; it is not held-out evidence. +- Finite saved render/depth/calibration arrays at iterations0/7000/10000; scalar history covers every step1–10000. +- Reference training completed in240.24s; review-stage wall time69.95s includes any shared-lane wait. +- [Summary](appearance_train_v1/review/summary.json), [agent findings and exact inspection coverage](appearance_train_v1/review/agent_findings.json), [run flags](appearance_train_v1/run_record.json). + +The four final sheets, loss curves and eight rotation frames were visually inspected. Full video playback is provided for owner review; sampled inspection is not certification of every orbit frame. + +Appearance acceptance is separate from the accepted prior and does not authorize a physical fit. + +Owner decision, 2026-10-29: “Accept this appearance with the disclosed limitations”. The [acceptance record](../owner_appearance_acceptance_c6a_v1.json) names the exact reviewed run and final PLY. No retraining or physical fitting is authorized. diff --git a/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/PRIOR_REVIEW_c6a_v1.md b/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/PRIOR_REVIEW_c6a_v1.md new file mode 100644 index 0000000000000000000000000000000000000000..e45f3c191321922fd64be362fe523570ce4db218 --- /dev/null +++ b/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/PRIOR_REVIEW_c6a_v1.md @@ -0,0 +1,29 @@ +# 225908 prior review — accepted 2026-10-28 + +The owner accepted this exact prior on 2026-10-28; [literal approval and bound identities](../owner_prior_acceptance_c6a_v1.json) are saved. [Trained appearance review](APPEARANCE_REVIEW_c6a_v1.md) is separate. + +This is the current C2 orange-belt prior for your accepted **328–560** packet. The initial reference images are converted frame **328** / native-container frame **329** in all four cameras, selected nearest to the initial 30 Hz sample. [Exact timestamps and local indices](prior_review_v1/source_frame_mapping.json) are retained. + +The fit follows the visible ring in all four views. The hidden arc behind the hand is inferred by the closed spline. It stopped after **one 25-iteration round** at the prescribed gradient tolerance, with length **639.923 mm**, maximum curvature **18.291/m**, and unchanged **2.95 mm tube radius**. Geometry checks passed: one watertight component, no self-intersection. + +| Camera | Visible centreline samples inside object mask | Native-image overlays | +|---|---|---| +| 12540311 | 99.66% | [centreline](prior_orange_v1/review/spline/centerline_0_with_masks.png) · [tube](prior_orange_v1/review/tube/silhouette_0_unmasked.png) | +| 12933256 | 99.71% | [centreline](prior_orange_v1/review/spline/centerline_1_with_masks.png) · [tube](prior_orange_v1/review/tube/silhouette_1_unmasked.png) | +| 17736741 | 98.05% | [centreline](prior_orange_v1/review/spline/centerline_2_with_masks.png) · [tube](prior_orange_v1/review/tube/silhouette_2_unmasked.png) | +| 23007103 | 100.00% | [centreline](prior_orange_v1/review/spline/centerline_3_with_masks.png) · [tube](prior_orange_v1/review/tube/silhouette_3_unmasked.png) | + +Review the [rotation video](prior_orange_v1/review/tube/orbit.mp4), [eight rotation samples](prior_review_v1/orbit_samples.png), and [geometry against the depth cloud](prior_orange_v1/review/tube/tube_scale_views.png). + +Depth uncertainty remains visible: per-camera median point-to-centreline residuals are **6.11 / 4.48 / 8.21 / 6.62 mm**. The lowest tube vertex is **0.996 mm below the table plane**; no clamp or vertical correction was applied. These residuals are distances to the centreline, not measured surface error. The observed depth clouds have camera-dependent vertical spread. + +The appearance initializer contains **16,144 observed + 51,840 mesh seeds = 67,984** points. Of 864 mesh vertices, 590 have observed colours and 274 receive the reference median-colour fill. These are appearance seeds, not final mechanical particle counts. Static appearance training was pending at the initial prior review and has since completed; see the separate appearance review above. + +Accepting this prior admits its actual mesh/manifest for the reference mechanical input construction and static Gaussian training. It does not accept a future trained appearance or authorize physical fitting. Your C6a instructions require this separate prior review. + +[Agent inspection record](prior_review_v1/agent_findings.json) · [Fit metrics](prior_orange_v1/steps/fit.json) · [Geometry checks](prior_orange_v1/tube/tube_checks.json) · [Manifest](prior_orange_v1/manifest.json) + +Owner-accepted identities: + +- Prior manifest SHA256: `03150af1696414ab83bc4544d13ccb51b42e42327e55a3d16a17d97bf91b483b` +- Tube GLB SHA256: `f967bddaa129d459e54be51335b8019460c824236b3cdad262dd6e3b01f03dfd` diff --git a/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_check_v1.log b/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_check_v1.log new file mode 100644 index 0000000000000000000000000000000000000000..b41b7f692870149efe25c227a2c23d8e0eea9ce0 --- /dev/null +++ b/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_check_v1.log @@ -0,0 +1,57 @@ +Found metadata.json file, assuming customized QQTT dataset! [28/10 08:41:49] +Init points from pcd... [28/10 08:41:50] +Init points from mesh... [28/10 08:41:50] +Loading Training Cameras [28/10 08:41:50] +Loading Test Cameras [28/10 08:41:50] +Number of points at initialisation : 67984 [28/10 08:41:50] +Sampling 100000 points from mesh [28/10 08:41:50] +{ + "cameras": [ + { + "id": 0, + "width": 1920, + "height": 1080, + "object_px": 8107, + "occlusion_px": 194245, + "valid_depth_px": 1955408, + "boundary_px": 3319, + "projection_max_error_px": 2.1358435451237037e-05 + }, + { + "id": 1, + "width": 1920, + "height": 1080, + "object_px": 7757, + "occlusion_px": 154853, + "valid_depth_px": 2025790, + "boundary_px": 3230, + "projection_max_error_px": 1.500418492931131e-07 + }, + { + "id": 2, + "width": 1920, + "height": 1080, + "object_px": 8011, + "occlusion_px": 208124, + "valid_depth_px": 1982692, + "boundary_px": 3291, + "projection_max_error_px": 4.743846685825791e-06 + }, + { + "id": 3, + "width": 1920, + "height": 1080, + "object_px": 5295, + "occlusion_px": 163967, + "valid_depth_px": 1906299, + "boundary_px": 2513, + "projection_max_error_px": 3.156711982019261e-06 + } + ], + "hybrid": { + "vertex_color_branch": true, + "mesh": 51840, + "observed": 16144, + "total": 67984 + } +} [28/10 08:41:51] diff --git a/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_check_v1_argv.json b/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_check_v1_argv.json new file mode 100644 index 0000000000000000000000000000000000000000..fe173f8baa91a20fe5447c6099e3bbcbe98cf036 --- /dev/null +++ b/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_check_v1_argv.json @@ -0,0 +1,18 @@ +{ + "argv": [ + "/mnt/data0/magna/belt_perception_wt_c6a_lab_input_preparation_20261027/scripts/real_rgbd_gs_train_env/.venv/bin/python", + "scripts/real_rgbd_appearance_train.py", + "check", + "--prior", + "/mnt/data2/magna/belt_perception_shared/phystwin_model_corpus_expansion/lab/session_20260921_225908_522335/window_anchor0328-0560/accepted_input_preparation_c6a_v1/prior_orange_v1", + "--accepted-prior-manifest-sha256", + "03150af1696414ab83bc4544d13ccb51b42e42327e55a3d16a17d97bf91b483b", + "--output", + "/mnt/data2/magna/belt_perception_shared/phystwin_model_corpus_expansion/lab/session_20260921_225908_522335/window_anchor0328-0560/accepted_input_preparation_c6a_v1/appearance_check_v1" + ], + "cwd": "/mnt/data0/magna/belt_perception_wt_c6a_lab_input_preparation_20261027", + "environment": { + "PYTHONDONTWRITEBYTECODE": "1", + "MPLBACKEND": "Agg" + } +} diff --git a/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_check_v1_result.json b/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_check_v1_result.json new file mode 100644 index 0000000000000000000000000000000000000000..0dfae8df8ea9836a07a3dcee2beb8c0acb406008 --- /dev/null +++ b/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_check_v1_result.json @@ -0,0 +1,4 @@ +{ + "returncode": 0, + "elapsed_s": 3.9478547510007047 +} diff --git a/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_review_v1.log b/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_review_v1.log new file mode 100644 index 0000000000000000000000000000000000000000..5106671a693c5b2bf60253ed4e82e9a32f815fea --- /dev/null +++ b/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_review_v1.log @@ -0,0 +1,32 @@ +/mnt/data0/magna/belt_perception_wt_c6a_lab_input_preparation_20261027/scripts/real_rgbd_gs_train_env/.venv/lib/python3.10/site-packages/torch/utils/cpp_extension.py:1965: UserWarning: TORCH_CUDA_ARCH_LIST is not set, all archs for visible cards are included for compilation. +If this is not desired, please set os.environ['TORCH_CUDA_ARCH_LIST']. + warnings.warn( +Found metadata.json file, assuming customized QQTT dataset! [28/10 08:46:49] +Init points from pcd... [28/10 08:46:50] +Init points from mesh... [28/10 08:46:50] +Loading Training Cameras [28/10 08:46:50] +Loading Test Cameras [28/10 08:46:50] +Number of points at initialisation : 67984 [28/10 08:46:50] +Sampling 100000 points from mesh [28/10 08:46:50] +Loading trained model at iteration 7000 [28/10 08:46:52] [28/10 08:46:52] +Found metadata.json file, assuming customized QQTT dataset! [28/10 08:46:52] [28/10 08:46:52] +Init points from pcd... [28/10 08:46:52] [28/10 08:46:52] +Init points from mesh... [28/10 08:46:52] [28/10 08:46:52] +Loading Training Cameras [28/10 08:46:52] [28/10 08:46:52] +Loading Test Cameras [28/10 08:46:52] [28/10 08:46:52] +Sampling 100000 points from mesh [28/10 08:46:52] [28/10 08:46:52] +Loading trained model at iteration 10000 [28/10 08:46:54] [28/10 08:46:54] [28/10 08:46:54] +Found metadata.json file, assuming customized QQTT dataset! [28/10 08:46:54] [28/10 08:46:54] [28/10 08:46:54] +Init points from pcd... [28/10 08:46:54] [28/10 08:46:54] [28/10 08:46:54] +Init points from mesh... [28/10 08:46:54] [28/10 08:46:54] [28/10 08:46:54] +Loading Training Cameras [28/10 08:46:54] [28/10 08:46:54] [28/10 08:46:54] +Loading Test Cameras [28/10 08:46:54] [28/10 08:46:54] [28/10 08:46:54] +Sampling 100000 points from mesh [28/10 08:46:55] [28/10 08:46:55] [28/10 08:46:55] +{ + "snapshots": [ + "0", + "7000", + "10000" + ], + "review": "/mnt/data2/magna/belt_perception_shared/phystwin_model_corpus_expansion/lab/session_20260921_225908_522335/window_anchor0328-0560/accepted_input_preparation_c6a_v1/appearance_train_v1/review" +} [28/10 08:47:01] [28/10 08:47:01] [28/10 08:47:01] diff --git a/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_review_v1_argv.json b/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_review_v1_argv.json new file mode 100644 index 0000000000000000000000000000000000000000..12c2ad23ce61c26f50a263c162aa7050988223a7 --- /dev/null +++ b/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_review_v1_argv.json @@ -0,0 +1,13 @@ +{ + "argv": [ + "/mnt/data0/magna/belt_perception_wt_c6a_lab_input_preparation_20261027/scripts/real_rgbd_gs_train_env/.venv/bin/python", + "scripts/real_rgbd_appearance_review.py", + "--run", + "/mnt/data2/magna/belt_perception_shared/phystwin_model_corpus_expansion/lab/session_20260921_225908_522335/window_anchor0328-0560/accepted_input_preparation_c6a_v1/appearance_train_v1" + ], + "cwd": "/mnt/data0/magna/belt_perception_wt_c6a_lab_input_preparation_20261027", + "environment": { + "PYTHONDONTWRITEBYTECODE": "1", + "MPLBACKEND": "Agg" + } +} diff --git a/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_review_v1_result.json b/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_review_v1_result.json new file mode 100644 index 0000000000000000000000000000000000000000..9fe7ddbb64bf11baf7ed34392d6c60af4b2a9f68 --- /dev/null +++ b/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_review_v1_result.json @@ -0,0 +1,4 @@ +{ + "returncode": 0, + "elapsed_s": 69.94856720499956 +} diff --git a/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_train_v1.log b/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_train_v1.log new file mode 100644 index 0000000000000000000000000000000000000000..302795ddb4e9aa3710791fecbe1fc097900a7b07 --- /dev/null +++ b/sessions/session_20260921_225908_522335/preprocessing/c6a-input-preparation-v1/native/appearance_train_v1.log @@ -0,0 +1,19 @@ +Optimizing /mnt/data2/magna/belt_perception_shared/phystwin_model_corpus_expansion/lab/session_20260921_225908_522335/window_anchor0328-0560/accepted_input_preparation_c6a_v1/appearance_train_v1/model +Output folder: /mnt/data2/magna/belt_perception_shared/phystwin_model_corpus_expansion/lab/session_20260921_225908_522335/window_anchor0328-0560/accepted_input_preparation_c6a_v1/appearance_train_v1/model [28/10 08:41:53] +Found metadata.json file, assuming customized QQTT dataset! [28/10 08:41:53] +Init points from pcd... [28/10 08:41:53] +Init points from mesh... [28/10 08:41:53] +Loading Training Cameras [28/10 08:41:53] +Loading Test Cameras [28/10 08:41:53] +Number of points at initialisation : 67984 [28/10 08:41:53] +Sampling 100000 points from mesh [28/10 08:41:53] + Training progress: 0%| | 0/10000 [00:00