APS-VLA: Action-Guided and Future-Aware Token Pruning for Efficient Vision-Language-Action Models
Summary
- Researchers introduce APS-VLA, an action-guided and future-aware visual token pruning framework designed to accelerate vision-language-action (VLA) models in robotic manipulation.
- A lightweight Decision Influence Estimator scores tokens using current visual observations, language instructions, and robot proprioception, trained via action supervision and an auxiliary visual transition objective.
- The framework keeps the pretrained VLA backbone frozen and removes the auxiliary predictor at deployment after performing a single hard token selection and packing step.
- Experiments show APS-VLA reduces mean policy inference latency from 137.25 to 93.25 ms (a 1.47x speedup) while maintaining task success rates across benchmark tasks.
Why it matters
APS-VLA provides a modular, plug-in method to reduce inference latency in vision-language-action models without retraining backbone policy parameters or degrading task performance.