In deployment of the VLA models to real-world robotic tasks, execution speed matters. In previous work arXiv:2510.26742 we analyze how to make neural computation of VLAs on GPU fast. However, we leave the question of how to actually deploy the VLA system on the real robots open. In this report we describe a set of practical techniques to achieve the end-to-end result of running a VLA-driven robot at an impressive speed in real world tasks that require both accuracy and dexterity. The stack of technology ranges across calibration, planning & control, and learning based method to identify optimal execution speed. In the tasks we show, the robot even executes in a speed on par with casual human operation and approaching the hardware limit of our lightweight arm. The unaccelerated videos and inference traces are provided in https://dexmal.github.io/realtime-vla-v2/.
Figures & tables
Figure 1: Example tasks used in this paper. See https://dexmal.github.io/realtime-vla-v2/ for unaccelerated videos.
Figure 2: Time delay in the system and train-inference discrepancy.
Figure 3: Effect of pre-amplification of robot command. We have to increase the magnitude of change in the robot command so that the actual position tracks model’s target.
Figure 4: Post-processing framework of VLA’s trajectory.
Figure 5: Distribution of failure timestamps during autonomous rollouts at 1× , 2× , and 3× execution speeds. At each speed, failures cluster around a distinct bottleneck stage, revealing a small set of phases that dominate the overall failure rate.
Figure 6: Human-in-the-loop speed modulation during closed-loop rollout. A human operator provides real-time “throttle” inputs to switch the robot between different execution speeds, slowing it down in precision-critical phases and accelerating it during low-risk motion. The plot shows joint-velocity traces from a real rollout; blue-shaded intervals indicate execution at 3× speed, while red-shaded intervals indicate operator-triggered slowdowns to 1× . This procedure provides a practical way to collect speed-modulation supervision directly from real-world execution.
Figure 7: Relationship between delay time and max inference speed.
Figure 8: Velocity, acceleration and jerk after and before post-processing.