Skip to content

armgit / moss

Public Imported from GitHub
0

MOSS: a small, mostly 3D-printed litter-picking rover. Cleaner streets, one maker at a time.

Python
4.0 KB
Raw Blame

Geometry provenance and validation boundary

R92 release geometry

The solid components were tessellated from the published MOSS_Printable_BRep_Parts_R92.step. Its SHA256 is 85e663ca8b7b784560de1649d18552fe11d97ee142986222d70d54dcbe573aeb. The source was checked against GitHub's release-asset digest on 26 October 2026.

STEP assembly placements were retained. OpenCascade/CadQuery tessellation used 0.15 mm linear and 0.15 rad angular deflection settings. These are mesh generation settings, not measured manufacturing or simulation accuracy. The configuration selects common parts, the Jetson carrier and the RealSense insert. Alternative Pi/Gemini and split-print assemblies are not superimposed on the model.

The eight printed SO101 shells use their original indexed geometry, matched against the R92 STL volume and assembly bounds. The correspondence to the print orientation is recorded in reference/print_to_rig.json: scale the published STL from millimetres to metres, then apply that rigid transform. One degenerate triangle in the base motor holder is omitted. Registration errors are numerical mesh correspondence checks, not hardware calibration results.

Six existing SO101 housing meshes retain their upstream Apache-2.0 provenance. The mounted belts come from the saved R92 assembly reference. Both retain all 513,216 original triangles, in OBJ format to avoid MuJoCo's binary-STL face-count limit; no voxel remeshing or decimation is used. Their centre separation is 270 mm, and their minimum Z establishes the ground origin.

Individual assets, input filenames/hashes, licences, rig assignments, triangle counts and world-space bounds are recorded in model/manifest.json. Runtime requires only the files included here, not the original Fusion installation, CAD export folders, or older simulation package.

Arm and gripper

Joint definitions originate from MuJoCo Menagerie SO101, commit 0059d4335f8156206f63a35662313385f7ad6d74. The MOSS R31 collision-review snapshot supplied its arm/tool rigid groups, NC90 slide/pinion relation and decomposed colliders. The generator input in reference/arm_skeleton.xml retains the joints/defaults and explicit contact exclusions, without the old chassis geometry. R92 placement is applied in reference/kinematics.json.

Old moving-part colliders are reused only where the R92 geometry matches their rig-local volume within 1% and world bounds within 0.5 mm; individual values are recorded. This establishes correspondence for the reused approximation, not an error bound for its contact surfaces. Other components use current per-part convex hulls. The R92 bin is decomposed separately so its interior stays open.

The upstream arm gains/inertias and the estimated chassis/tool parameters have not been fitted to MOSS measurements. The gripper's preliminary damping/friction and position control replace values that had only been used for static pose inspection. The current package is not Dirac Robotics' calibrated model.

Reproduction scope

tools/build_model.py deterministically rebuilds both MJCF files from the shipped assets and reference data. No external service or API is used. Re-tessellating the manufacturing STEP requires an OpenCascade environment and preserving the source assembly transforms, configuration selection and rig frames above. Regenerating collisions can change results and requires re-running the tests.

Validation checks both stored asset hashes and the compiled MuJoCo mesh transforms against the source assembly bounds. The checks also exercise stationary integration, forward motion, turning, heading-relative motion, small arm movements, coupled jaw motion, the bin cavity and a released object settling inside it. These are offline software checks, not an autonomous pickup trial or a measurement of real-world accuracy.