The pebble motion problem (PMP) asks whether one configuration of labeled pebbles on a graph can be transformed into another by moving pebbles to adjacent unoccupied vertices. It is a fundamental model of graph reconfiguration and is closely related to multi-agent path finding (MAPF). A central open problem since Kornhauser, Miller, and Spirakis (FOCS 1984) is to understand the origin of the classical $Θ(N^3)$ worst-case behavior. While it is known that every feasible instance on an $N$-vertex graph admits a solution sequence of length $\Ord(N^3)$, it has remained unclear which instances actually require cubic complexity. In this paper, we resolve the long-standing complexity of the pebble motion problem on trees. We show that every feasible instance on an $N$-vertex tree admits a solution sequence of length $\Ord(N^2 \log N)$, computable by an output-sensitive algorithm. Since a lower bound of $Ω(N^2)$ is known, this establishes that the $Θ(N^3)$ phenomenon does not occur on trees and nearly closes the gap $Ω(N^2)\le \OPT(N)\le \Ord(N^3)$ up to a logarithmic factor. Building on this result, we extend our approach to general graphs by applying the tree algorithm to breadth-first spanning trees. This yields an efficient framework that produces $o(N^3)$-length solution sequences for a broad class of instances, including the classical square-grid example, where we recover the $\Ord(N^{3/2})$ bound observed by Kornhauser, Miller, and Spirakis. Finally, by analyzing the behavior of this algorithm, we obtain strong structural restrictions governing when $Θ(N^3)$ complexity can arise. We show that such behavior is possible only under highly constrained conditions, specifically when $Θ(N)$ degree-two vertices lie on cycles of length $Θ(N)$, with each cycle being the shortest containing the corresponding vertex.