---
title: "Booth 208 Was Booked. Nobody Came."
date: 2026-11-17
category: Humanoids & Robotics
site: NeuroAI
canonical: https://neuroai.site/a/na-robot-gemini-robotics-mercedes-1117
language: en
---

# Booth 208 Was Booked. Nobody Came.

> The conference app listed OpenAI at booth 208. The floor plan agreed.

The conference app listed OpenAI at booth 208. The floor plan agreed. A Fortune reporter found an empty rectangle: no robot, no signage.

The booth was paid for. The exhibitor did not show.

That same week at Mercedes-Benz's Sindelfingen plant, a humanoid lifted metal brackets from a bin and laid them into a sequencing tray in body order. DeepMind scored the motion at 96%.

The scenes sit three days apart: Gemini Robotics 2.5's blog on November 10, Fortune's report on OpenAI's robotics team on November 13. Chinese outlets called it "same day," which is wrong.

## An empty booth is not an exit

OpenAI split a team of a few dozen. Algorithm researchers went to the main model group under chief research officer Mark Chen; applied robotics and hardware went to the Applied division under president Greg Brockman. The line was not shut. The company stopped believing robots need a brain separate from the main model.

The premise is GPT-6 Astra, shipped in September, which already handles some robot work without dedicated training.

Fortune's account: early in the year the plan was still the Sora route, teaching a video model how objects move. Months later the general model went around it. The pivot was expensive, since world simulation was the project's original justification.

OpenAI closed a robotics lab in 2020, rebuilt it in 2024 with Figure, and split again in February 2025. Six years, two dissolutions, opposite reasons: first the models could not learn the task, now they do not need to.

Altman's humanoid program continues, targeting a first public demo in 2027. The booth is empty. The project is not.

## Whose 96% is it

Apptronik's Apollo 2 does the work, with a SharpaWave hand: five fingers, 22 degrees of freedom. The task is kitting, pulling parts from bulk bins and placing them in build order before the main line, so a worker can reach without looking.

The 96% belongs to no single model. Gemini Robotics ER runs the process; 2.5 runs the arm and hand.

You cannot buy "a 96% model." You buy a system: two models, one body, one hand.

The same model runs on Franka Duo's two-finger gripper, so a new body needs no retraining: cross-embodiment generalization, in DeepMind's phrase.

What DeepMind left out: trials, duration, robot count, failure distribution. Without them, 96% proves a concept, not a line you can hand over.

## 24 right, and the 25th wrong

96% is 24 out of 25.

Fine on a launch screen, terrible in quality control, where 4% is 40,000 PPM.

But kitting is not quality control. A missed grasp breaks nothing. Reach again.

So 96% gates people, not parts. Three assumptions not in DeepMind's blog: 300 grasps per station per hour, 12 seconds each, 40 seconds of human time to clear a failure. At 96%, that is 12 failures an hour, one every 5 minutes, or 8 minutes of labor per station per hour, 13% of a work hour.

At 30% of a worker's time, he watches 2.25 stations.

The price of 96%: one robot, plus 0.44 of a person.

## The threshold sits at 98.5%

To push that person past 6 stations, success must reach 98.5%.

Run it backwards: six stations, 300 grasps each per hour, 40 seconds per intervention, labor under 30%. Mean time between failures must stretch to 13 minutes, 65 grasps with one error allowed.

64 of 65. 98.46%.

The gap is 2.5 points. What it buys is 2 stations per person versus 6.

The threshold is a division, not a standard: acceptable labor share divided by intervention time over mean time between failures. Below 3, the robot has not replaced the worker. It has moved him from picking parts to picking up the parts the robot dropped. A factory never buys "can it grasp." It buys "how many can one person watch."

A shift is 2,400 grasps and 96 failures. Whoever absorbs those 96 is the real cost line.

## Two bets, not one

The disagreement is not whether to build robots. It is where robot intelligence comes from.

OpenAI: from general models, at scale. No dedicated team, just a bigger model.

DeepMind: from specialized models in two layers. One runs the process and may be slow; one runs hands and feet and must be fast. Between them, Gemini 4 Argon builds simulations and digital twins and watches failure video, keeping slow work in the cloud and fast work at the edge. The general model is not excluded, only demoted to a tool that makes training material.

OpenAI's evidence comes from Nvidia's Astranomy research: the RoboLab-120 benchmark, 120 pick-and-place tasks. Lakeside, a specialized model, scored 51.0%. Astra acting directly scored 66.3%. Codex with session history reached 74.8%.

That figure came with the simulation paused while the model thought. Inference averaged 18.56 seconds, at about $36.40 in API fees per success.

A conveyor belt will not wait 18 seconds.

DeepMind's number came off a real line, sample undisclosed. The denominators differ, so ranking 74.8% against 96% was the week's least meaningful comparison in Chinese tech media.

## Who stands at 208 next year

Neither side gave the number that decides it: a full shift, no line stop.

Not 96 against 74.8. Who first puts "8 hours, 2,400 grasps, no more than 2 interventions" on paper. That needs better than 99.9%, 3.9 points away. At DeepMind's stated pace of "clear improvement after a few weeks of training," not out of reach, and not next launch either.

The positions do not conflict. OpenAI is half right: a robot needs no separate brain. DeepMind is half right: it does need dedicated hands.

Brains can be general. Hands cannot.

Booth 208 will probably have someone next year. What stands there may be a table of human interventions per shift.

*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-gemini-robotics-mercedes-1117
Free to quote with attribution and a link to the original.
