The Signal · Issue 01 · Physical Intelligence
The humanoid hype gap, measured
Reports say Tesla is ramping Optimus production, while Google DeepMind's latest robotics model controls whole humanoid bodies. The distance between production targets and reliable work is the number to track.
Facts are attributed and linked to their sources. The analysis and opinions are the author's own.
What it is
TechRepublic, citing The Information, reports that Tesla increased Optimus production at its Fremont factory from a few dozen robots a week in the second quarter of 2026 to several hundred in August, and that managers are targeting more than 1,000 a week by the end of 2026. Tesla has not confirmed these figures. Elon Musk has said public sales could begin by late 2027, provided the robots meet expectations for reliability, safety and functionality, and Tesla has not announced final pricing or a launch date.
For context on the models that drive such machines: on 30 July, Google DeepMind introduced Gemini Robotics 2, a vision-language-action model it says can control full humanoids "from feet to fingertips", demonstrated on Apptronik's Apollo 2 walking to a table, picking up a watering can and placing it on a shelf. DeepMind notes its robots "have more to advance in movement speed".
Why it matters
Production volume and useful capability are different metrics, and humanoids are where they diverge most. A robot built is not a robot doing reliable work. Unconfirmed production figures, conditional sales dates and demos with stated limitations are three different kinds of evidence, and they should not be added together.
A robot built is not a robot doing reliable work. Count the second number.
The responsible reading is to separate them: what has been confirmed, what has been reported, what has been demonstrated, and under what conditions.
Production ramps also create their own measurement risk. When a factory target is set in robots per week, it is tempting to count robots shipped rather than robots doing reliable work, the same trap as counting dashboards built rather than decisions improved. The healthier pairing is output plus a field-performance metric reported alongside it.
I think of this as an evidence ladder. At the bottom are targets and ambitions. Then reported figures from credible outlets, labelled as unconfirmed. Then company-published demos with stated limitations. Then company-published evaluations with fixed rubrics and held-out conditions. At the top, independent replication. Most humanoid news sits on the lower rungs, and it is fine to report it, provided the rung is labelled. This magazine tries to do that consistently.
The measure I would most like to see standardised is interventions per hour of autonomous work, alongside full-task success. It is the robotics equivalent of an error rate in production analytics, and it is what a buyer would actually pay for. The interactive autonomy ladder in this section lets you play with how intervention rates and supervision interact, with stylised numbers.
None of this is pessimism. The underlying models are improving quickly, and companies are publishing more rigorous evaluations. The point is simply that the best decisions come from reading each claim at its true strength.
What I'd watch
- Confirmed deployment metrics: hours of autonomous work, interventions per hour, task success under fixed rubrics.
- Safety standards for humanoids working near people, and who certifies them.
- Whether makers publish evaluation protocols as detailed as the best ones this season.
Sources
- Tesla Optimus timeline: robots targeted for 2027 launch. TechRepublic, Oct 2026.
- Gemini Robotics 2 brings whole body intelligence to robots. Google DeepMind, 30 Jul 2026.