---
title: "A Robot Needs a Command Every 12 Milliseconds. Chat Cannot Deliver That."
date: 2026-11-19
category: Humanoids & Robotics
site: NeuroAI
canonical: https://neuroai.site/a/na-robot-embodied-brain-1119
language: en
---

# A Robot Needs a Command Every 12 Milliseconds. Chat Cannot Deliver That.

> Eighty-three divided by 0.4 is 207.5. Both come from Anthropic's frontier red team report.

Eighty-three divided by 0.4 is 207.5.

Both come from Anthropic's frontier red team report. Low-level control for a legged robot needs about 83Hz. A general model in non-reasoning mode runs at 0.2 to 0.4Hz.

A gap of 200 to 400 times. Not an opinion. A physical constraint.

## Twelve milliseconds versus 2.5 seconds

83Hz means one command every 12 milliseconds. 0.4Hz means one every 2.5 seconds. At 0.2Hz, one every five.

In those 12 milliseconds the loop reads the sensor, computes joint torque, sends a command and waits. There is room to calculate. No room to think.

A general model is strong elsewhere. It breaks a sentence into steps: "sort the parts into boxes" becomes walk over, bend, identify, grasp, place. That is planning, and 2.5 seconds is generous.

So the question is not whether a general model works. It is which layer belongs to it.

At one meter per second, a model checked every 2.5 seconds has moved 2.5 meters before it looks again; at 0.2Hz, five. A low-level loop still catches the fall, but the gap is two orders of magnitude.

That loop is the last moat. A cloud call must survive a network round trip, a queue, inference and the return path. Twelve milliseconds is gone before that. Control has to sit on a local board with tight power and thermal budgets. Physics decides the location.

## The planning layer is already being eaten

On November 13, OpenAI confirmed it had dissolved its independent robotics research team.

The same week, GPT-6 Astra drove an open-source arm costing about $150 through a cloud API. With no fine-tuning, it completed work such as drawing from an image.

A $150 arm plus a general model, doing real work. That is uncomfortable for anyone selling a robot brain.

Apple's HALO prototype took the same route. A Unitree G1 humanoid carrying a Mac mini reached 91% success on a single task in six real homes with no pre-built map.

An off-the-shelf body, a desktop computer, a general model.

This is what absorption looks like. Task understanding, step decomposition and retry once meant state machines and rule libraries. Now they go to a model that has seen enough scenes. The cost is that it cannot say why it failed. One wrong step poisons the rest.

The caveat is the sample: 91% on one task, in six homes. It proves the thing runs, not that it ships.

## Three routes, one missing layer

The news crowded into one week. On November 9, Anthropic released Claude Fable 5.2, at 64.9% average success on RoboDojo simulation. On the same board, GPT-6 Astra scored 29.8%, Gemini 4 Argon 39.4%, Opus 5.5 34.2% and Fable 5.1 27.6%. The leader was 25.5 points ahead of second and 35.1 ahead of Astra.

The next day, Google DeepMind released Gemini Robotics 2.5, at 96% task success on a Mercedes-Benz line, with the Apollo 2 humanoid sorting parts on its own.

On November 16, Huawei Cloud open-sourced CloudRobo-WM1.2 and took first overall on RoboTwin 2.0 and LiberoPro 2.0. That week Nvidia released GR00T 2 at CoRL 2026, and Ant Lingbo open-sourced LingBot-Touch, a tactile foundation model.

Touch deserves a note. Vision says where the cup is; touch says how hard to squeeze. That call happens at contact, where planning cannot reach, so a tactile model lives in the control layer.

Five companies, five angles on the same missing layer.

Their scores are high. None is playing at the 83Hz layer.

64.9% is simulation. 96% is a line. First place is a leaderboard. All three measure whether a task finished, not whether a command lands every 12 milliseconds. And they do not share a denominator: modeled simulation, a fixed station, peers scoring peers.

Shen Yujun of Ant Lingbo says it plainly: evaluation has no consensus, and model, data and deployment are coupled. Bad simulation proves a model is bad. Good simulation does not prove a real machine works.

## The business becomes control

If planning is absorbed, what remains for a dedicated brain company? The control layer.

It is not attractive. It tells no story, only latency, jitter and robustness.

But it is hard to replace. A general model is called request by request. Control must be always on, on a fixed cadence, where one missed tick is an incident. More parameters do not mean less latency. Jitter, which is fast then slow, is worse.

Tian Feng argues general models will take planning and the pretraining base, not real-time control. The capability threshold comes in one to two years, commercial scale later, and the real inflection when evaluation gets cheap.

Nobody knows whether a robot model works. First place in simulation can fall on a real machine. 96% on one line can fail at the next station. Proving it costs space, bodies, people and time, and thicker data makes the next version steadier. The entrance to that loop is evaluation, not an algorithm.

So the business shifts from selling models to selling control plus a data loop: control by unit, data by scene, neither of which a general model supplies.

## Convergence to three to five

Zhu Xing is blunter: China's embodied sector will not support many physical foundation model players. A first consolidation comes before next year; three to five remain long term.

Three to five is a narrow door. Anthropic, Google, Huawei, Nvidia and Ant Lingbo are already through it. Most robot brain companies become integrators, or get absorbed.

The capability survives. The standalone company does not.

Body makers know each unit's jitter, latency and wear. Model companies have seen more scenes. How those advantages split the bill is the real negotiation.

One question left. If the inflection is evaluation cost, will the first company to turn real-machine validation into an assembly line be worth more than the first to reach 64.9%?

*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-robot-embodied-brain-1119
Free to quote with attribution and a link to the original.
