Agentic AI
VOTE: Ensemble Voting Technique Cuts Robot Vision-Language Model Inference Time by 39x
Researchers propose a training and inference optimization framework that dramatically reduces latency in vision-language-action models for robotic manipulation while improving task performance.
A new research effort tackles the computational bottlenecks that plague large-scale vision-language-action (VLA) models used to control robots. The approach, called VOTE, combines two key innovations: a compact training framework that minimizes token generation and an ensemble voting technique that extracts more value from each action prediction.
The Problem with Current VLA Models#
Modern VLA systems—which interpret images and natural language instructions to generate robot commands—suffer from two significant limitations. First, they generate excessive tokens during inference, creating prohibitive latency that makes real-time robotic control impractical on edge hardware. Second, the actions these models produce often aren't fully exploited, wasting computational effort and leaving performance gains on the table.
The VOTE Solution#
The framework tackles both issues simultaneously. Researchers developed a training approach that reduces the number of action tokens produced, enabling parallel processing to cut inference costs and training expenses. This efficiency gain carries real-world impact: the system achieves 39 times faster inference than OpenVLA (a widely-used baseline), reaching 46 Hz throughput—fast enough for actual robotic deployment on edge platforms.
Beyond raw speed, VOTE introduces a novel voting-based ensemble strategy that combines current and historical action predictions. This mechanism improves how well generated actions translate into task success, effectively wringing more utility from each model output.
Results and Practical Impact#
The approach outperforms state-of-the-art VLA baselines on robotic manipulation benchmarks, posting substantially higher success rates alongside the dramatic inference speedup. The combination of speed and accuracy makes the system genuinely deployable in real-world robotic settings, addressing a major gap between academic progress and practical robotics applications. Source code is available online.
