Vision-language-action (VLA) policies can solve the same manipulation task through different action interfaces, yet task success alone does not guarantee that their physical executions agree. We examined cross‑policy end‑effector geometry using 15,000 closed‑loop LIBERO rollouts from four policies.
Under clean conditions we formed 3,600 configuration‑matched (hence dependent) policy pairs. Pairs where both policies succeed have a median normalized dynamic time warping (DTW) distance of $0.0120\ \text{m}$, whereas pairs with exactly one success show $0.0380\ \text{m}$. This ordering holds across every task, every policy pair, and nine sampling and band‑limited representations; however, the ratio varies severalfold across representations, so we report only the direction rather than a fixed multiple.
Both‑failure pairs are even more separated but rest on thin, uneven support, and are therefore reported as exploratory.
Within successful executions, swapping partners creates larger geometric differences across tasks than across initial states.
A matched baseline still reveals measurable, heterogeneous residual policy differences, indicating that a low cross‑policy distance does not imply interchangeability.
Successful executions are about as far from same‑task demonstrations as those demonstrations are from each other, suggesting task‑associated geometry without cleanly separating training‑data overlap from task constraints.
A common 72‑action window preserves the ordering but reduces its magnitude; adjusting endpoints and duration likewise leaves a positive mixed‑outcome coefficient relative to both‑success pairs, though its magnitude depends on the specification.
Under composite visual stress, policy rankings and pair composition change together.
Review