Vision‑language‑action (VLA) policies exhibit strong generalization in robotic manipulation, yet dense visual tokens cause high inference latency. Visual token pruning reduces this cost, but the challenge lies in selecting a compact token set that still yields reliable actions. Existing approaches rely mainly on static or immediate heuristics. APS‑VLA learns token selection from robot demonstrations while keeping the pretrained VLA backbone frozen. A lightweight Decision Influence Estimator scores tokens conditioned on the current visual observation, language instruction, and robot proprioception. Training combines two complementary signals: action supervision forces the selected tokens to reproduce the full‑input policy predictions and match demonstrated actions; an auxiliary visual transition objective uses demonstrated actions to predict future visual features from the selected representation. At deployment, the estimator performs a single hard selection and packing step before the policy’s native action‑generation, and the auxiliary predictor is removed. Experiments with OpenVLA‑OFT on LIBERO and with $\pi_{0.5}$ and XR‑1 on RoboCasa365 show inference acceleration with only modest changes in task success. In five real‑world dual‑arm tasks, mean policy inference latency drops from 137.25 ms to 93.25 ms, a $1.47\times$ speedup, while success rate rises from 60% to 61%. These results demonstrate that APS‑VLA serves as a modular plug‑in framework that improves efficiency across VLA architectures without altering underlying policy parameters or the native action‑generation pipeline.
Review: APS‑VLA achieves notable inference speed‑up through learned token pruning while preserving, and even slightly improving, task success, making it a practical enhancement for real‑world robotic control systems.