The robot no longer needs a brain of its own

Physical Intelligence, one of the companies that train models for robots, opened an API on 13 October 2026 under a title that sums up the sector's bet: «Frontier Models for Every Robot» ^[1]. A customer uploads its robot's data, fine-tunes the company's model there and runs it from the cloud, without building a machine-learning team of its own ^[1]. A day earlier, a Harvard group had published an experiment that goes further. It put GPT-6-Astra, a general OpenAI model that nobody tuned for the task, in charge of a simulated robot arm, and the model beat the best system trained for that benchmark, with 92.1 % success against 60.8 % ^[2].

The week is best read as an instance of something larger. The robot's brain is coming apart from its body and becoming a component that is rented and swapped. A general model already does much of the work that used to require training a system for every machine, provided someone gives it an interface to look and move with, and that interface is the piece that stays when the model changes. If that holds, what limits the deployment of robots at scale stops being training and becomes the body, the data of the real operation and the work of integration. The region was never going to train those models, and what is now scarce is precisely what it has at hand.

A model nobody trained to move an arm

The Harvard benchmark, RoboMME, gathers 16 manipulation tasks that demand temporal, spatial, object and procedural memory ^[2]. The model sees the whole episode and dictates every position of the arm's end effector and every command to the gripper, without a single adjustment of its parameters for those tasks ^[2]. It succeeded in 737 of 800 episodes, above the 90.5 % that people had achieved with a privileged interface, and matched or beat those people on 13 of the 16 tasks ^[2]. Each episode cost about 39 US cents ^[2].

A team of six researchers had shown the same idea in July with ordinary coding agents, and updated its work on 7 October ^[3]. Their system, VIA, gives the agent a virtual gripper that it moves with a handful of tools, and only sends it to the robot once it is satisfied with the result ^[3]. They tested Claude Code and Codex unmodified, with six different models, and success grew with the size of the model: 38 %, 73 % and 100 % for the three tiers of Codex in simulation ^[3]. Outside simulation, Codex with Astra 6 completed 24 of 25 attempts with a real mobile robot ^[3]. The authors put it bluntly: a coding agent is, in a sense, secretly a robot-control agent ^[3].

What the specialist sells

Physical Intelligence does not rent out a general model but a robotics model, which the customer fine-tunes with its own demonstrations ^[1]. Its first users include Actor, whose wheel loader fills a dump truck in a real commercial deployment with nobody in the cab, and Almond, which obtained in four days a model to deburr parts of its own robot ^[1]. NEURA Robotics presented Neuron AI on 14 October, a single model for five different bodies made by the company ^[4]. It runs on the robot itself without depending on the cloud and has no published results yet ^[5].

Google DeepMind had announced a similar architecture under its own name on 30 July ^[6]. Its reasoning model, Gemini Robotics ER 2, serves as the robot's «high-level brain» and is offered in Google AI Studio, while a second, smaller model runs on the machine ^[6]. That second model adapts to a new two-armed robot with a few hours of data, typically fewer than 200 examples ^[6]. Figure broke off its agreement with OpenAI in February 2025, and its founder wrote at the time that language models had become, for his company, «the smallest piece of the puzzle» ^[7]. Twenty months later, that piece tops a manipulation benchmark without task-specific training.

Three layers and an interface

A paper from the University of California, San Diego, and Nvidia, published on 2 October, describes the division of labour that is emerging ^[8]. Frontier models can already manipulate objects from a few demonstrations, but their latency is too high to control a robot in real time, so the general model gives the instructions and a small, fast local policy drives the motors ^[8]. That combination succeeded in 65 % of attempts, against 51.25 % for the local policy alone ^[8].

A study presented this year at ICML, the main machine-learning conference, marked the limit from the other side ^[9]. The twelve models it tested perform better the more human-designed scaffolding they receive and worse when it is taken away, although a training-free agent that iterates and reviews its own results approaches human level on several tasks ^[9]. What remains is a three-layer architecture: a general model that plans and corrects itself, a local policy that acts, and between them an interface of legible tools. The interface is the only layer that does not have to be rebuilt when a better model comes out.

The short leg of the buildout

Robotics is still the smaller part of the investment in artificial intelligence. Nvidia's chief financial officer estimated in August that the five largest cloud providers will spend nearly 800 billion dollars in 2026 and 1.3 trillion in 2027 ^[10]. The company's physical AI business, the one of self-driving cars and robots, passed 9 billion in twelve months, while its data centres billed 75 billion in a single quarter ^[11]. Its founder spoke of that segment as a hope for the next five years ^[11]. Factories installed more than 600,000 industrial robots in 2025, and the International Federation of Robotics expects 655,000 this year ^[12].

Venture capital is running ahead of sales. Humanoid companies raised 11 billion dollars so far in 2026, more than double the record of all of 2025, according to Dealroom ^[13]. Five robotics companies raised 510 million in the week of 12 October to build houses, assist in surgery, manufacture drugs and pack orders ^[13]. One of them, Ultra, leases its warehouse robots by the month and runs them on Physical Intelligence's software ^[13].

The body, the other half

Chinese manufacturers shipped more than 97 % of the world's humanoids in the first half of 2026, about 19,100 units, according to the count TechNode cites ^[14]. AgiBot shipped about 8,400 and Unitree about 5,900 ^[14]. Unitree's pages list its R1 humanoid today from 4,900 dollars and the G1 from 13,500, excluding tax and shipping ^[15]. Faraday Future launched in the United States on 15 October a 20-joint humanoid for developers at 2,490 dollars ^[16].

The United States has already shut the door on the next generation of those bodies. On 28 July the Federal Communications Commission added to its list of risky equipment every «advanced robotic device» produced outside the country, except those granted a conditional approval by the Department of War ^[17]. The definition covers humanoids, quadrupeds and autonomous mobile robots weighing more than two kilos, and models already authorised can still be sold ^[18]. The national-security determination warns that such robots could be commandeered remotely ^[17]. The risk has a precedent: in 2025 researchers disclosed a flaw in the Bluetooth of several Unitree robots that gave full control over them and could jump from one to another ^[19]. China's Ministry of Commerce called the measure «unilateral bullying», according to CNN ^[20].

What still fails today

Each call to the model in the Harvard experiment took 18.3 seconds on average, and the authors warn that their result measures a capability, not a real-time controller ^[2]. The tasks are simulated, on a tabletop and with a fixed camera ^[2]. VIA only works for near-static tasks and could not catch a ball ^[3]. Physical Intelligence acknowledges that some of its first deployments from the cloud ran into connectivity problems ^[1].

Until recently, putting a robot to work demanded a brain made to measure, with its own machine-learning team and months of tuning for every machine. That bill kept advanced robotics for the companies that could pay it. The robot's intelligence becomes a service that is contracted and replaced, and every new model released for coding or for using a computer reaches the robot at no extra cost, because the interface is the same. The VIA authors say it plainly: improvements in agent perception, reasoning and tool use transfer directly to robot control ^[3]. That changes where robotics sits in the buildout: a robot with a rented brain does not wait for infrastructure of its own, it plugs into the data centres already being paid for and buys inference by the episode like any other customer ^{[2] [10]}.

What becomes possible in the region is a work cell with no artificial intelligence lab behind it. An integrator in Monterrey or Campinas can combine an off-the-shelf Chinese body with a model rented by use, and contribute the interface and the tests. At the official exchange rate of 16 October, an R1 costs about 24,900 reais before duties and shipping ^{[15] [21]}. That is a little more than a year of Brazilian minimum wage, counting the thirteenth salary ^[22]. The tasks that fit best are those that tolerate the robot thinking for a few seconds: inspecting parts, sorting returns, preparing orders, moving samples in a laboratory. Brazil is already installing industrial robots at an unprecedented pace, up 38 % in 2025 to nearly 4,300 units, driven by investment from Chinese carmakers ^[12]. Mexico chained its third year of decline and fell below 5,200 ^[12], so its mid-sized plant, which could never pay for a classic automation project, is the natural customer for that cell.

The body is the other half of the calculation. The United States no longer authorises new models of mobile robots made outside its territory ^[17], and we found no equivalent rule in Latin America. That leaves the region as the large open market where the best American head, rented through an API, can legally be mounted on the cheapest Chinese body. Mercado Libre shows the region does not yet occupy that place: its humanoid is a Digit from the American company Agility working in Texas, and its Latin American warehouses appear as cases to explore ^[23].

The cycle Physical Intelligence describes makes clear where the value lies: collect demonstrations, train, test and use the failures to decide what to collect next ^[1]. That cycle belongs to whoever operates the task, not to whoever makes the model, and the region has the tasks in the mine, the field, the warehouse and the industrial kitchen. It is the lesson of Nº 014 seen from the workshop floor: the asset is the check, the set of real cases that says whether the machine got it right.

The counterpoint carries weight. None of this week's results is a robot in production, because the best one is a simulation with a fixed camera that thinks for 18 seconds per step ^[2], and the specialist selling the brain still charges to tune it with the customer's own data ^[1]. A model without dated versions can change under the robot without warning ^[2], and that clashes with any certification that requires knowing what software runs on the machine. The cloud is not guaranteed either in a mine or a field with patchy signal, which is where the local models of Google and NEURA earn their place ^{[6] [5]}. A cheap body with a known Bluetooth flaw is a real risk inside a plant ^[19], and Washington could ask its partners for what it already demands of itself. We found not a single robot in the region governed by a general model, so the opportunity is described and not yet demonstrated.

THE PLAY

1. Build the interface, not the brain. Make legible tools to look, point and move that any model can operate, and switch models when a better one comes out without touching the robot.
2. Start with the tasks that can wait. Inspection, sorting and the laboratory tolerate seconds of thought; leave to a local policy whatever cannot wait or lose signal.
3. Write the tests first. The asset is the set of your customer's real episodes and what must come out of each, because it is how you will judge any head you rent tomorrow.