Driving · 2026-11-11
APS-VLA: Action-Guided and Future-Aware Token Pruning for Efficient Vision-Language-Action Models
Xiaoyang Xing, Chun-Mei Feng, Anthony Cohn, Xiaoliang Gong, Asoke Nandi
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
APS-VLA is a framework for vision-language-action (VLA) token pruning that learns token selection from robot demonstrations while keeping a pretrained VLA backbone frozen. It trains a lightweight Decision Influence Estimator (DIE) using action supervision and an auxiliary action-conditioned visual transition prediction. At deployment, DIE performs one hard token selection/packing step before the frozen policy’s native action generation, enabling faster inference with small task-success changes.
Why it matters
Processing dense visual token sequences can dominate VLA inference latency. APS-VLA addresses the challenge of learning which visual tokens remain control-relevant across task stages by using demonstration trajectories and a future-aware auxiliary objective, while making deployment non-invasive (frozen backbone; auxiliary head removed). The reported real-world results show reduced policy latency with comparable success, supporting modular plug-in efficiency for multiple VLA architectures.
Method
- Learn a budgeted visual token selector (DIE) from robot demonstrations while freezing the pretrained VLA backbone.
- Train DIE with joint supervision: action supervision that matches both dense full-context policy predictions and demonstrated actions, plus action-conditioned visual transition prediction from the selected representation.
- Deploy by scoring once and performing a single hard selection/packing step per camera; the auxiliary predictor is removed and the frozen policy runs its native action-generation procedure.
Limitation
The paper does not state any specific limitation on APS-VLA beyond reporting performance trade-offs with higher pruning ratios (e.g., stronger pruning leading to sharper success drops).
Abstract (from arXiv)
Vision-language-action (VLA) policies have demonstrated strong generalization in robotic manipulation, yet processing dense visual tokens often incurs high inference latency. Visual token pruning offers a practical way to reduce this cost, but identifying a compact set of tokens that preserves reliable action generation remains challenging. Existing methods largely rely on static or immediate heuristic cues to estimate token importance. We introduce APS-VLA, a framework that learns visual token selection from robot demonstrations while keeping the pretrained VLA backbone frozen. A lightweight Decision Influence Estimator scores tokens conditioned on current visual observations, language instructions, and robot proprioception. Its training combines two complementary signals: action supervision matches both the full-input policy predictions and demonstrated actions, while an auxiliary visual transition objective uses demonstrated actions to predict subsequent 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 procedure; the auxiliary predictor is removed. Experiments with OpenVLA-OFT on LIBERO and with $\pi_{0.5}$ and XR-1 on RoboCasa365 demonstrate inference acceleration with modest changes in task success. Across five real-world tasks on two dual-arm platforms, APS-VLA reduces mean policy inference latency from 137.25 to 93.25 ms, achieving a $1.47\times$ speedup with an average success rate of 61%, compared with 60% for the unpruned baseline. Together, these results support APS-VLA as a modular, plug-in framework for efficient robotic control across VLA architectures, preserving the underlying policy parameters and native action-generation procedure.
Related papers
- PaceVLA: Adaptive Temporal Action Quantization in Vision-Language-Action Models
- Fast-VLA: You Only Think Once for Robot Manipulation
- Causal Analysis of Robot Foundation Models Enables Efficiency Gains
- WAVE-0: Bridging VLA Policies and WAMs with Lightweight Latent Future Prediction
- ViDAL: A Visual Dynamics-Grounded Action Latent Space for Vision-Language-Action Models
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First announced by arXiv: 2026-11-11 · Indexed by Pub-AI: 2026-11-11 UTC
APS-VLA: Action-Guided and Future-Aware Token Pruning for Efficient Vision-Language-Action Models
First announced by arXiv: 2026-11-11 · Original record ↗
- arxiv
- 2611.13030
Explicit source versions (1)
- v1 · Date not supplied
- Preprint available
Source date: 2026-11-11 · First detected: 2026-11-11 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2611.13030v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv
Vision-language-action (VLA) policies have demonstrated strong generalization in robotic manipulation, yet processing dense visual tokens often incurs high inference latency. Visual token pruning offers a practical way to reduce this cost, but identifying a compact set of tokens that preserves reliable action generation remains challenging. Existing methods largely rely on static or immediate heuristic cues to estimate token importance. We introduce APS-VLA, a framework that learns visual token selection from robot demonstrations while keeping the pretrained VLA backbone frozen. A lightweight Decision Influence Estimator scores tokens conditioned on current visual observations, language instructions, and robot proprioception. Its training combines two complementary signals: action supervision matches both the full-input policy predictions and demonstrated actions, while an auxiliary visual transition objective uses demonstrated actions to predict subsequent 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 procedure; the auxiliary predictor is removed. Experiments with OpenVLA-OFT on LIBERO and with $\pi_{0.5}$ and XR-1 on RoboCasa365 demonstrate inference acceleration with modest changes in task success. Across five real-world tasks on two dual-arm platforms, APS-VLA reduces mean policy inference latency from 137.25 to 93.25 ms, achieving a $1.47\times$ speedup with an average success rate of 61%, compared with 60% for the unpruned baseline. Together, these results support APS-VLA as a modular, plug-in framework for efficient robotic control across VLA architectures, preserving the underlying policy parameters and native action-generation procedure.
