Latency and Hardware Matching of Vision-Language-Action Models for Humanoid Robots

Abstract

Vision-Language-Action (VLA) frameworks empower robots to execute complex tasks from natural language instructions, but mismatched inference speed and low-level control frequency limit reliable industrial and daily deployment. A VLA model may select a reasonable action, but a robot still has to receive it in time. This review examines the problem through RoboNurse-VLA, VLA-Pilot, and Sliding-Cache VLA. RoboNurse-VLA generates commands at 5 Hz, while its Universal Robot 5 (UR5) controller runs at 30 Hz. The faster controller updates the arm between model outputs, but it cannot make a late command safe. VLA-Pilot takes 2.41 seconds for one steering step. This may be acceptable in a stable supervised task, but an observation can become outdated when a person or object moves. In one OpenVLA test, Sliding-Cache VLA reduced inference time from 52.4 to 34.1 milliseconds. This was an improvement in that experiment, not a direct advantage over the other systems. These cases show that model accuracy alone is not enough for physical deployment. Response time, low-level control, and separate safety tests must also be considered.