SpecPrune-VLA: Accelerating Vision-Language-Action Models via Action-Aware Self-Speculative Pruning
SpecPrune-VLA: Accelerating Vision-Language-Action Models via Action-Aware Self-Speculative Pruning
This research presents SpecPrune-VLA, a training-free two-level pruning method for Vision-Language-Action (VLA) models. The approach leverages spatial-temporal consistency in VLA tasks by combining local information with global context to reduce visual tokens. It features action-level static pruning using global history and local attention, layer-level dynamic pruning based on importance, and a lightweight action-aware controller that adjusts pruning aggressiveness according to end-effector speed. Experiments demonstrate up to 1.57x speedup in LIBERO simulation and 1.70x on real-world tasks with negligible impact on success rates.
The paper introduces SpecPrune-VLA, a training-free pruning method for Vision-Language-Action models that combines local attention with global context to achieve up to 1.70x speedup on real-world tasks while maintaining negligible success rate degradation.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
speedup factor
≤1.57
View original evidence
chieves up to 1.57$\times$ speedup in LIBERO simulation and 1.70$\times$ on real-world tasks, with negligible success rate degradation.
Open source S5speedup factor
1.7
View original evidence
chieves up to 1.57$\times$ speedup in LIBERO simulation and 1.70$\times$ on real-world tasks, with negligible success rate degradation.
Open source S5Reported success rate
>20%
View original evidence
arXiv:2509.05614v4 Announce Type: replace-cross Abstract: Pruning is a typical acceleration technique for compute-bound models by removing computation on unimportant values. Recently, it has been applied to accelerate Vision-Language-Action (VLA) model inference. However, existing acceleration methods focus on local information from the current action step and ignore the global context, leading to >20% success rate d
Open source S2
- paper: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
chieves up to 1.57$\times$ speedup in LIBERO simulation and 1.70$\times$ on real-world tasks, with negligible success rate degradation.
Open source S5
arXiv:2509.05614v4 Announce Type: replace-cross Abstract: Pruning is a typical acceleration technique for compute-bound models by removing computation on unimportant values. Recently, it has been applied to accelerate Vision-Language-Action (VLA) model inference. However, existing acceleration methods focus on local information from the current action step and ignore the global context, leading to >20% success rate d
Open source S2
Source：arXiv Robotics — research abstracts · arxiv.org