Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
id
int32
0
26
label
int64
0
26
mesh.vertices
listlengths
31.1k
51.2k
mesh.faces
listlengths
62.3k
102k
mean_width
float32
0.07
0.08
diameter
float32
0.09
0.09
roundness
float32
0.37
0.45
aspect_ratio
float32
0.71
1
volume
float32
0
0
concavity
float32
1
1
0
0
[[-0.02681674063205719,-0.026891404762864113,-0.02687367983162403],[-0.02681674063205719,-0.02689140(...TRUNCATED)
[[13971,14519,556],[28814,28823,27009],[24902,24901,24903],[24903,24901,24895],[22849,22850,26921],[(...TRUNCATED)
0.078751
0.090001
0.422211
0.973903
0.000161
1
1
1
[[-0.02662353403866291,-0.0262486319988966,-0.026379860937595367],[-0.02662353403866291,-0.026248631(...TRUNCATED)
[[1,163,477],[17176,33861,17177],[13521,13522,17174],[29773,33960,29772],[37183,16936,16937],[16935,(...TRUNCATED)
0.077394
0.090004
0.406408
0.906669
0.000155
1
2
2
[[-0.026620853692293167,-0.026257408782839775,-0.02673051878809929],[-0.026620853692293167,-0.026257(...TRUNCATED)
[[10197,400,556],[240,162,10196],[29884,13524,13523],[10204,400,10201],[400,10199,10201],[400,10197,(...TRUNCATED)
0.077393
0.090003
0.406365
0.890922
0.000155
1
3
3
[[-0.026642169803380966,-0.02657686173915863,-0.02639899030327797],[-0.026642169803380966,-0.0265768(...TRUNCATED)
[[32116,23263,32115],[477,13631,13603],[13603,13631,23263],[13603,32116,16968],[23263,32116,13603],[(...TRUNCATED)
0.077563
0.090003
0.406924
0.844579
0.000155
1
4
4
[[-0.02664078399538994,-0.026577986776828766,-0.017486220225691795],[-0.02664078399538994,-0.0265779(...TRUNCATED)
[[6851,82,6537],[27637,10297,27773],[27773,27636,27637],[27637,6536,10297],[17333,333,17343],[6539,1(...TRUNCATED)
0.077564
0.090004
0.406891
0.844332
0.000155
1
5
5
[[-0.02712458185851574,-0.02656751126050949,-0.02686859667301178],[-0.02712458185851574,-0.026567511(...TRUNCATED)
[[1,84,398],[32588,20235,32587],[4884,3878,4852],[20107,3878,4884],[8,3060,9],[32586,32585,34820],[3(...TRUNCATED)
0.078243
0.090001
0.421349
0.757604
0.000161
1
6
6
[[-0.026802152395248413,-0.026428937911987305,-0.026371978223323822],[-0.026802152395248413,-0.02642(...TRUNCATED)
[[37099,26404,26405],[26405,37098,37099],[10284,32857,10283],[36874,20188,36876],[36876,20188,10280](...TRUNCATED)
0.077313
0.090004
0.406457
0.941992
0.000155
1
7
7
[[-0.026801809668540955,-0.026428598910570145,-0.017511241137981415],[-0.026801809668540955,-0.02642(...TRUNCATED)
[[36970,36971,20186],[35067,32817,32816],[84,398,1],[38634,37054,38623],[28018,32890,32862],[20259,3(...TRUNCATED)
0.077312
0.090003
0.406458
0.941408
0.000155
1
8
8
[[-0.026801809668540955,-0.017390968278050423,-0.026548869907855988],[-0.026801809668540955,-0.01739(...TRUNCATED)
[[3,400,635],[38767,37060,38763],[13278,13279,20187],[36874,35252,20191],[20190,36874,20191],[20191,(...TRUNCATED)
0.077312
0.090003
0.406457
0.779719
0.000155
1
9
9
[[-0.026772432029247284,-0.026407847180962563,-0.01719314232468605],[-0.026772432029247284,-0.026407(...TRUNCATED)
[[15456,6636,6637],[242,6639,1],[13276,319,241],[6638,15456,6637],[18146,18187,18219],[18146,18106,1(...TRUNCATED)
0.07714
0.090001
0.405942
0.922668
0.000155
1
End of preview. Expand in Data Studio

Filleted cubes, four edges of twelve

27 classes. Each class is a cube whose twelve edges carry one of two fillet radii: four edges at 0.36 of the half-edge (9.75 mm on the 54 mm cube that fits a 9 cm caliper diameter), eight at 0.02 (0.54 mm). The fillets run the full length of every edge and blend at the corners, so the surface is smooth everywhere and the solid is convex -- exactly, volume/hull-volume 1.000000, which keeps it on the loader's exact-hull path and away from VHACD.

A rolling-ball fillet of a convex solid is convex at every radius. OCCT does not quite deliver that at this radius: where two large fillets meet a small one its mixed-radius vertex blend dips below the ideal surface, costing 0.17% of volume and putting the as-built solid at 0.9983, under the loader's threshold. The hull is the intended shape and the dip is the artifact, so the meshes shipped here are hulled. The audit does not move either way.

The 27 classes are the orbits of the four-edge subsets of a cube under the cube's rotation group, so no rigid rotation carries one class onto another. Nine of them are chiral pairs; reflections are not quotiented out, because a mirror image is a genuinely different object to a hand.

Episodes randomise scale on a grid of 50 rungs, so absolute size carries nothing and what has to be controlled is the scale-invariant shape. All classes share one largest caliper width to 0.004%, and reading a scale-invariant descriptor to 2 percent and taking the nearest class gives: worst inertia eigenvalue/d^2 4.28x chance, volume/d^3 4.09x, area/d^2 3.65x, sphericity 1.11x. For comparison the set this replaces, abc20cyl, was solved to 0.775 against 0.05 chance -- 15.5x -- by height-over-diameter alone.

This is the deliberately deeper point on the relief/leak frontier: at r_big 0.18 the same construction audits at 3.2x but carries only 1.76 mm of relief contrast, which is marginal for an Allegro fingertip read through joint angles. Here the contrast is 3.81 mm and the leak is 4.3x. Fillet relief sets both quantities and they track each other at correlation 0.988, so on a cube the trade cannot be avoided, only chosen.

Meshes are watertight single bodies in metres, scaled so the largest caliper width is 0.090 m. id is the class index.

Downloads last month
57