Expand synthetic generator: geometric diversity #8

New Issue

forbes · 2026-02-02T19:33:21Z

forbes commented

2026-02-02 19:33:21 +00:00

Summary

Add geometric diversity to synthetic assembly generation to ensure the model trains on realistic geometric configurations. Per Phase 1.2 of the repository plan.

Features to implement

Random orientations and positions

Bodies placed with random 3D positions (not just along a line)
Random orientations via scipy.spatial.transform.Rotation
Anchor points derived from body geometry, not always at midpoints

Parallel axis detection

Intentionally generate assemblies with parallel joint axes
These trigger geometric degeneracies that the Jacobian catches but the pebble game misses
Critical for training the model to recognize when combinatorial analysis is insufficient
Configurable ratio of parallel-axis assemblies in a batch

Grounded vs floating variants

Generate both grounded (one body fixed to world) and floating (free-floating) assemblies
Configurable ratio in batch generation
Grounded assemblies have 0 trivial DOF; floating have 6

Axis alignment diversity

Joint axes sampled from: cardinal axes, random unit vectors, near-parallel perturbations
Ensures the model does not overfit to axis-aligned configurations

Requirements

RandomPoseGenerator helper (or methods on SyntheticAssemblyGenerator) for diverse body placement
Parallel axis injection with configurable probability
grounded parameter on all template generators
Joint axis sampling strategies: cardinal, random, near-parallel
Update batch generation config to control geometric diversity parameters

Depends on

#5 (generator port)
#7 (assembly templates)

## Summary Add geometric diversity to synthetic assembly generation to ensure the model trains on realistic geometric configurations. Per Phase 1.2 of the repository plan. ## Features to implement ### Random orientations and positions - Bodies placed with random 3D positions (not just along a line) - Random orientations via `scipy.spatial.transform.Rotation` - Anchor points derived from body geometry, not always at midpoints ### Parallel axis detection - Intentionally generate assemblies with parallel joint axes - These trigger geometric degeneracies that the Jacobian catches but the pebble game misses - Critical for training the model to recognize when combinatorial analysis is insufficient - Configurable ratio of parallel-axis assemblies in a batch ### Grounded vs floating variants - Generate both grounded (one body fixed to world) and floating (free-floating) assemblies - Configurable ratio in batch generation - Grounded assemblies have 0 trivial DOF; floating have 6 ### Axis alignment diversity - Joint axes sampled from: cardinal axes, random unit vectors, near-parallel perturbations - Ensures the model does not overfit to axis-aligned configurations ## Requirements - [ ] `RandomPoseGenerator` helper (or methods on `SyntheticAssemblyGenerator`) for diverse body placement - [ ] Parallel axis injection with configurable probability - [ ] `grounded` parameter on all template generators - [ ] Joint axis sampling strategies: cardinal, random, near-parallel - [ ] Update batch generation config to control geometric diversity parameters ## Depends on - #5 (generator port) - #7 (assembly templates)

forbes added the phase:1 feature labels 2026-02-02 19:33:21 +00:00

forbes referenced this issue

2026-02-02 19:33:53 +00:00

Dataset generation CLI #10

forbes referenced this issue from a commit

2026-02-02 20:57:56 +00:00

feat: geometric diversity for synthetic assembly generation

forbes closed this issue

2026-02-02 20:57:56 +00:00

forbes referenced this issue from a commit

2026-02-03 16:54:16 +00:00

Merge pull request #8 from Ondsel-Development/name_change