Persistent DOFs. Evolving coordination.
Same physical DOFs. Evolving relations. Changing relation orders.
Robot DOFs are persistent, finite physical entities, while their statistical coordination changes with temporal context and task. Some Square windows contain higher-order residuals beyond a pairwise maximum-entropy model, motivating a Dynamic Hypergraph representation that can naturally reduce to a graph.
Significant-window poster · MP4
Left: one representative Square demonstration. Cursor: its normalized time. Shaded band and relation snapshot: the aligned 200-demo cohort window. Background quarters are time guides, not semantic motion phases; significance is a population-level result.
A: the same DOF pair changes correlation sign across normalized windows. B: the same Panda has different task-specific relations. C: after accounting for all three pairwise marginals, Square has 54 FDR-significant triplet × windows; Lift, Can and ToolHang have no FDR-significant triplet-window detected under the current test. D: Square's higher-order rankings remain consistent across splits, bin counts and window sizes.
Panda · 7 arm DOFs · 20 Hz · 200 proficient-human demonstrations per task. Residual tests use one window-mean sample per demonstration and BH q < 0.05 across all triplet × window tests within each task. Associations motivate a representation; they do not establish causality.
This repository presents HyperDOF’s research motivation and preliminary statistical evidence. Implementation code and datasets are not included.
The demonstrations and visualization environment build on robomimic and robosuite.