Robots maintain control speed as their AI models grow tenfold

Separating slow visual reasoning from fast action updates preserved 20 updates per second across model sizes and improved humanoid pick-and-place success.

Top University

Bingzhang Wang · Nathan Zhao · Jialeng Ni · Jiajin Zheng · Yin Zhang · Kehua Chen · +1 more

XPENG Robotics · University of Washington

Research Digest·11 November 2026·2 min read

Illustration: Robots maintain control speed as their AI models grow tenfold

In one line

Decoupling the vision-language backbone from the action head lets humanoid control rate stay at 20 Hz regardless of model size.

Before you read this

Three ideas this paper builds on

vision-language-action modelA robot policy model that maps visual observations and language instructions to actions, trained on demonstrations and often used for manipulation tasks.83 papers →role decouplingSeparating distinct functions such as planning, reasoning, and execution into separate components improves efficiency, quality, and reliability.48 papers →flow matchingA generative training method that learns an ordinary differential equation mapping noise to data by matching velocity fields along probability paths.167 papers →

Read it as

Controlling a humanoid robot usually means the robot must wait while a large, powerful vision-language model decides what to do. This waiting time grows with the model's size, slowing down the robot and making its movements jittery or inaccurate.

DASH solves this by splitting the robot's brain into two parts that work at their own pace: a slow, thoughtful part that understands the scene, and a fast, reactive part that moves the hands. The fast part moves the robot 20 times per second, no matter how big the slow part is.

In tests, DASH kept a steady control rate of 20 Hz while the standard approach slowed from 9.7 Hz to 4.1 Hz as the model grew. On a real humanoid robot, DASH succeeded in 23 of 24 pick-and-place trials, versus 16 out of 24 for the standard pipeline.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

vision-language-action model role decoupling latency-agnostic scaling desynchronized training flow matching action chunking