Determine where the frame advantage decay stops

Determine where the decline in the motion-planning frame’s performance advantage beside local geometry stops as the training dataset grows.

Background

The paper reports that the canonical start-goal frame contributes a +22.4 percentage-point advantage over the world-frame representation when trained on 60 environments, but only +17.2 points when trained on 250 environments under a matched budget. This observed decay suggests that the benefit of the frame may narrow as more training data becomes available. The authors explicitly leave unresolved the point at which this decline stops, cautioning against extrapolating the measured margin to much larger datasets.

References

The frame's value beside local geometry falls with data. It is $+22.4$ points at $60$ environments and $+17.2$ at $250$ on a matched budget. We have not found where that decay stops, and a reader extrapolating to much larger datasets should expect the margin to narrow rather than hold.

— What Symmetry Buys a Learned Motion Planner  (2609.10033 - Sevincel, 9 Sep 2026) in Section Limitations