Add parameterized cube primitive benchmarks

[Yuan-Xinyi]

Stack

• Layer: 3/3
• Base: codex/composed-spatial-generalization
• Depends on: Add composed spatial-generalization benchmarks #779
• Followed by: none

Description

Adds the first-stage primitive-geometry generalization populations to the YAML-configured Atomic Skill benchmark.

The benchmark uses EmbodiChain's built-in cube primitive and exposes geometry through the existing objects[].size parameter. Each size is defined once as a YAML anchor; the runtime object override and the domain provenance reference the same value. A new primitive variation therefore has one parameter edit surface, while the resolved suite and every case manifest retain the exact object configuration.

Domain	Primitive	Parameter	Value in metres	Geometry change
nominal	cube	size	[0.05, 0.05, 0.05]	nominal baseline
cube_uniform_40mm	cube	size	[0.04, 0.04, 0.04]	uniform scale down
cube_uniform_60mm	cube	size	[0.06, 0.06, 0.06]	uniform scale up
cube_aspect_x	cube	size	[0.07, 0.042258, 0.042258]	X elongated
cube_aspect_y	cube	size	[0.042258, 0.07, 0.042258]	Y elongated
cube_aspect_z	cube	size	[0.042258, 0.042258, 0.07]	Z elongated

Primitive track	Atomic Skills	Populations
primitive-cube-rigid-object	PickUp, Pour, AxisAlign	nominal + five variants
primitive-cube-held-pose-target	MoveHeldObject	nominal + five variants
primitive-cube-place-target	Place	nominal + five variants
primitive-cube-position-target	PushObject	nominal + five variants

The stage is deliberately expressed through basic primitive parameters. Additional registered primitives and mesh-backed geometries can reuse the same domain structure in later stages. Primitive populations have independent track groups, so their summaries stay separate from position, orientation, workspace-shell, and composed-pose summaries.

This PR adds benchmark definitions and validation only; no benchmark results were run or committed.

Type of change

• Enhancement (non-breaking change which improves an existing functionality)

Screenshots

Not applicable.

Validation

• black . with black==26.3.1
• black --check --diff --color . — 5408 files unchanged
• python docs/scripts/check_api_docs.py — 2613/2613 exports documented
• python -m pytest -q -p no:cacheprovider -c /dev/null --noconftest tests/benchmark/motion_generation/test_contracts.py tests/benchmark/motion_generation/test_generalization.py — 27 passed
• Production loader and round-trip probe — 79 evaluation populations, 24 primitive populations, 4 primitive groups, round trip passed
• python -m py_compile tests/benchmark/motion_generation/test_atomic_pose_config.py
• Focused test_atomic_pose_config.py collection is unavailable in the current local environment because optional open3d is not installed

Checklist

• I have run the black . command to format the code base.
• I reviewed affected documentation and agent context. The benchmark README was updated; the context router reports no affected topics.
• Public API changes are reflected in the API docs; no public package API was added.
• I have added tests that prove primitive parameter isolation and full non-geometry configuration preservation.
• Dependencies have been updated, if applicable; no dependency changes were required.

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no actionable issues were found.

Summary

Adds five cube-size variants for six skills, with separate nominal populations across four tracks.

• Cube-based skills now compare five geometry variants with the 50 mm cube.

_{Reviews (2) · Last reviewed commit: "refactor(benchmark): parameterize cube p..." · Reviewed by Greptile}

[yangchen73]

Looks good.