VLA-RAIL: A Real-Time Asynchronous Inference Linker for VLA Models and Robots

TL;DR

VLA-RAIL tackles the real-time fusion gap between Vision-Language-Action models and robotic control by decoupling inference from execution in a plug-and-play client-server setup. It introduces a two-stage post-processing pipeline: intra-chunk smoothing with cubic polynomials and inter-chunk fusion via temporal alignment and dual-quintic spline blending, ensuring $\mathcal{C}^2$ continuity across chunks. The framework accelerates execution by jointly tuning trajectory interpolation and command dispatch frequencies, enabling faster-than-teleoperation performance while reducing motion jitter and increasing task success across diverse VLA models. Empirical results on real robots show smoother trajectories, faster task completion (up to $2.09\times$), and improved compatibility across multiple VLA architectures, highlighting its practical impact for scalable VLA deployment in manipulation tasks.

Abstract

Vision-Language-Action (VLA) models have achieved remarkable breakthroughs in robotics, with the action chunk playing a dominant role in these advances. Given the real-time and continuous nature of robotic motion control, the strategies for fusing a queue of successive action chunks have a profound impact on the overall performance of VLA models. Existing methods suffer from jitter, stalling, or even pauses in robotic action execution, which not only limits the achievable execution speed but also reduces the overall success rate of task completion. This paper introduces VLA-RAIL (A Real-Time Asynchronous Inference Linker), a novel framework designed to address these issues by conducting model inference and robot motion control asynchronously and guaranteeing smooth, continuous, and high-speed action execution. The core contributions of the paper are two fold: a Trajectory Smoother that effectively filters out the noise and jitter in the trajectory of one action chunk using polynomial fitting and a Chunk Fuser that seamlessly align the current executing trajectory and the newly arrived chunk, ensuring position, velocity, and acceleration continuity between two successive action chunks. We validate the effectiveness of VLA-RAIL on a benchmark of dynamic simulation tasks and several real-world manipulation tasks. Experimental results demonstrate that VLA-RAIL significantly reduces motion jitter, enhances execution speed, and improves task success rates, which will become a key infrastructure for the large-scale deployment of VLA models.

Figures (6)

• Figure 1: Illustration of current issue in VLA deployment (left) and the advantage of VLA-RAIL (right). Deploying VLA models on heterogeneous robots involves one-to-one adaptation, while VLA-RAIL provides a unified framework that enables plug-and-play deployment across diverse VLA models and heterogeneous robots.
• Figure 2: Temporal Pipeline Of Robotic Manipulation. There are three real-time pipelines happened concurrently during robotic manipulation execution. The first one, named eye pipeline, is the state retrieval pipeline that acquires the proprioception data and visual images via hardware drivers at a frequency of $f_{state}$. The second one, name brain pipeline, is the model inference pipeline that predicts new action chunks using current acquired state data at a frequency of $f_{infer}$. The third one, named hand pipeline, is the motion control pipeline that sends the action command to robot at a frequency of $f_{ctrl}$.
• Figure 3: Overview of VLA-RAIL. The asynchronous framework decouples VLA model inference with robotic motion control using a client-server architecture, with request–response communication implemented via ZMQ protocol. By exposing unified interfaces on both the server and client sides, VLA-RAIL supports plug-and-play deployment across multiple VLA models and heterogeneous robotic platforms. Functionally, the server hosts computationally intensive VLA model inference on GPU devices, while the client employs a multi-threaded design to concurrently handle proprioceptive and visual observation acquisition, model inference requests, trajectory post-processing, and real-time robot motion control.
• Figure 4: Trajectory smoothness comparison under three inference-execution strategies: without post-processing (Trajectory-1), naive switching (Trajectory-2), and VLA-RAIL (Trajectory-3). Left column: time-series plots of joint angle (top), velocity (middle), and acceleration (bottom). The zoomed insets highlight trajectory details, with the right inset directly comparing the original VLA output and VLA-RAIL post-processed result. Right column: standard deviation (std) of each metric computed at 1-second intervals, where lower values indicate smoother motion.
• Figure 5: Illustrations of the task completion speed under different inference–execution strategies. The bottle-grasping task (top) includes 3 stage: 1) a human points to a bottle while the robot recognizes and grasps it; 2) the robot delivers the bottle to the human; and 3) the human takes the bottle and the robot returns to its initial pose. The tea-pouring task (bottom) includes 3 stages: 1) recognizing and grasping the teapot handle; 2) adjusting the teapot pose above the cup; and 3) pouring tea and returning the teapot to its original position. The horizontal bars shows the average time spent in each stage and the total task duration. Here, VLA-RAIL achieves up to $2.09\times$ speedup compared to the strategy without any post-processing.