GPU acceleration has made motion planning one to two orders of magnitude faster, but each accelerated planner is built around a single paradigm—sampling, search, or optimization—so comparing across libraries also compares their collision checkers. PrismPlan addresses this by building the common bottleneck, batched collision detection and the batched forward kinematics beneath it, as a shared per-sphere device-code substrate that every planner consumes.
The authors ran six planners over 1,800 MotionBenchMaker problems on three robots (7–14 DoF, including a dual-arm whole-body composite) with an equal 2-second budget. Because all planners execute the same device code, the backend is held fixed by construction, enabling head-to-head paradigm comparison.
On the 7-DoF Panda, sampling (pRRTC) was the most reliable paradigm at 98%, while parallel search (wPA*SE) followed within four points at 3–4 times lower median latency. Local optimizers trailed well behind when made to solve from scratch. On the 14-DoF dual-arm composite, sampling still solved about 90% of feasible problems, while no search or optimization planner cleared a third of them.
The results suggest that sampling remains the most reliable paradigm, but parallel search can be competitive at lower latency when the backend is held constant. The open-source code is available with the paper.