Topology coverage
Robot training distribution
Which presets should a generalist controller train on? Hover a robot, or search for one by name, to find it in every plot. Click it to add it to your picks.
1Occupancy
- Shows
- Robots per node count and member type. Each cell reads in current split / in library.
- Look for
- Thin rows and empty or hatched cells that no split can cover.
Hatched: Henneberg graphs with no preset. “–”: not enumerated beyond 8 nodes.
2Topology map
- Shows
- Every robot placed by similarity (classical MDS); nearby robots are similar. The axes have no units.
- Look for
- Clusters, and regions that no training robot sits near.
3Descriptor matrix
- Shows
- Descriptor pairs, with Spearman ρ over all robots and within n = 8 above the diagonal.
- Look for
- Descriptors that still vary within n = 8: real axes beyond size. Drag in a panel to select robots; Esc clears.
Descriptor definitions
- Degree spread
- Standard deviation of node degree. Mean degree is fixed by n because every preset has 3n − 6 edges.
- Triangles / node
- Number of 3-cycles in the graph, divided by node count.
- Symmetry
- log₂ of the graph automorphism count.
- WCRI
- Initial worst-case rigidity index, on a log scale. Values below 10⁻⁶ (numerically flexible poses) are drawn at 10⁻⁶.
- Resting faces
- Convex-hull faces the robot can rest on stably (center of mass inside the face).
4Coverage
- Shows
- Distance from robots to their nearest training robot as the training set grows. n = 7 is never picked.
- Look for
- How far deliberate picks beat random, and where the curve flattens.
Farthest-point greedily minimizes the worst case, so it wins that metric partly by construction. Weighting changes only the mean.