One frontier model could soon steer a humanoid, a robot arm and a robotaxi. Here’s what that changes for the people building them.
A robot has always been an equation: a body, plus sensors, plus compute, plus an AI model trained for that one machine. The last few months have started to rewrite it. The model may be turning into a component that any robot can borrow, the way an AI agent already borrows a browser or a spreadsheet.
From robot-specific models to robot-use agents
MIT’s Phillip Isola gave the shift a name in a September essay: robot-use agents, where a language model drives a robot as a tool. His argument is that the old objections, weak spatial reasoning and thin physical intuition among them, are fading with newer models. Anthropic’s robotics research points the same way, with Claude controlling machines that range from a simple pendulum to a Unitree Go2 quadruped.
The tooling is arriving at the same pace. Waddle Labs launched an API in July that lets agents write and refine the control code for a customer’s robot, and says a working policy takes about 20 minutes. In August, Anthropic opened a research preview of the Model Hardware Standard (MHS), a shared specification that lets AI agents discover and operate physical devices such as robotic arms and liquid handlers. And when RoboCurve tested OpenAI’s GPT-6 Astra on a dual-arm setup, it placed a block in a bowl in 19 of 20 attempts.
Put those pieces together and a stack takes shape: frontier model, agent, hardware interface, skills and policies, machine. A humanoid, a robotic arm, a quadruped or an autonomous vehicle can all sit at the bottom of it.
What builders can stop building
If the intelligence can be borrowed, nobody needs to train a full AI stack before shipping a robot. Picture a start-up with a good idea for hospital logistics or sidewalk cleaning. Instead of hiring a team to build a model from scratch, it can spend its first year on the parts nobody can borrow:
- The embodiment, and the sensors and actuators that give it a body
- Edge compute that keeps the machine responsive on its own
- Physical-world data that nobody else holds
- Safety and reliability engineering
- The specific problem the machine exists to solve
A lower barrier means more products, more form factors and more fleets, working in more places. Something worth watching.
Early, and honest about it
The same Astra test is a useful reality check. On a precise puzzle-insertion task it succeeded in 2 of 20 attempts, and even the easy block task averaged around two and a half minutes per run. Isola himself concedes that language-model control still trails purpose-built solutions, and that latency and reliability hold it back from safety-critical work.
The builders are designing around that, not pretending otherwise. Under MHS, an agent can write a deterministic program once and let the device run it without the model thinking at every step, with safety limits enforced by the device driver rather than the model. Waddle keeps fixed platform primitives beneath the skills its agents generate. The pattern is consistent: the frontier model decides what to do, while specialised policies and edge inference decide how, in milliseconds.
What the robotaxi already shows
Autonomous vehicles are the nearest thing Physical AI has to a proven version of this architecture. A robotaxi doesn’t wait for a cloud model to decide when to brake. The driving stack runs on board, in real time, while the cloud handles everything around it: remote assistance when the vehicle meets something it can’t resolve, map and model updates for the whole fleet, and the data flowing back from every mile.
Isola’s essay approaches it from the other direction, suggesting a car could one day be as easy for a general model to steer as a factory arm. Whether or not that arrives soon, the commercial robotaxi already shows what the safe version looks like: onboard autonomy for the split second, and a connected layer for everything outside it.
Where connectivity fits
That split defines what the network has to do. The fast loop, balance, grip, braking, stays on the machine and needs no connection. The slow loop is where the cloud earns its place: planning a new task, writing a new skill, learning from what the whole fleet has seen. It runs in seconds rather than milliseconds, which makes it a question of reach and consistency more than raw speed.
Three flows depend on it. Fleet management keeps every machine visible, assigned and healthy. Over-the-air updates deliver a new skill, policy or model version to thousands of machines at once, which is how a fleet gets smarter without new hardware. And telemetry carries physical-world data back to the backend, feeding the next version.
How machines connect depends on where they work. Those that stay indoors, in warehouses, factories and hospitals, can rely on site Wi-Fi or a private cellular network. Those that work outdoors and move between sites, such as delivery robots, patrol quadrupeds, yard vehicles and robotaxis, have no site network to lean on, so cellular carries all three flows. Plenty of fleets span both, and the connectivity has to follow the machine rather than the building.
Scale is where this gets practical. A hundred robots in one warehouse is a networking project. Ten thousand across six countries is a provisioning, coverage and compliance project, and it decides how quickly a new skill reaches the whole fleet.
At Cubic3, we build automotive-grade IoT connectivity for fleets that move: multi-network coverage in every market, eUICC-based provisioning so a machine ships once and activates locally, and data residency controls that keep each market’s data where the law says it belongs.
Physical AI is early, and the architecture is still settling. The builders who plan now for how their machines will be updated, managed and heard from, wherever they work, will move fastest when it does.
Frequently asked questions
What is Physical AI?
Physical AI is artificial intelligence that senses and acts in the real world through a machine, such as a humanoid, a robotic arm, a quadruped or an autonomous vehicle. In the general-purpose approach, one frontier model can steer many different bodies, where each robot previously needed its own model trained for that one machine.
What are robot-use agents?
Robot-use agents are AI agents in which a language model drives a robot as a tool, the way an agent already uses a browser or a spreadsheet. MIT’s Phillip Isola named the idea in an essay published on 7 September 2026.
Can a large language model control a robot today?
Yes, with clear limits. In RoboCurve’s September 2026 test on a dual-arm rig, OpenAI’s GPT-6 Astra placed a block in a bowl in 19 of 20 attempts but completed a precise puzzle insertion in only 2 of 20, averaging around two and a half minutes per block run.
What is the Model Hardware Standard?
The Model Hardware Standard (MHS) is a shared specification from Anthropic that lets AI agents discover and operate physical devices such as robotic arms, liquid handlers and microscopes. Anthropic opened a research preview on 27 August 2026, and each device’s safety limits are enforced by its device driver.
How do Physical AI systems stay safe in real time?
The frontier model decides what to do, while specialised policies and edge inference on the machine decide how, in milliseconds. A robotaxi follows the same pattern: its driving stack runs on board, so braking never waits for a cloud model.
What connectivity does a Physical AI fleet need?
The fast loop of balance, grip and braking runs on the machine and needs no connection. The slow loop of fleet management, over-the-air updates and telemetry needs reliable reach across every site and market, which means cellular for machines that work outdoors or move between locations.
How does Cubic3 support robot and autonomous vehicle fleets?
Cubic3 provides automotive-grade IoT connectivity for fleets that move, with multi-network coverage in every market. eUICC-based provisioning lets a machine ship once and activate locally, and data residency controls keep each market’s data where the law says it belongs.
Plan your fleet’s connected layer now
See how Cubic3 keeps fleets that move updated, managed and reporting home
