Someone gave GPT-6 Astra a small robot and asked it to stand up. It stood up on its hands.
The experiment, shared informally online this week, is not a benchmark. As a snapshot of embodied AI's current state, though, it is hard to improve on: a general model can understand enough about a robot, its software and its goal to invent a coherent motor strategy, while still missing something as basic as which way up the body is supposed to be. (zamantika.com)
A more controlled experiment produced the same split. Robocurve gave Astra command of two real robot arms. On a simple block-into-bowl task, it succeeded in 19 of 20 attempts. On a precision insertion task, it managed only two. Astra was not controlling raw motor torques. It issued end-effector poses, inverse kinematics translated those commands into joint angles, its speed was capped at 25%, and conventional safety guardrails remained underneath it. (openai.robocurve.org)
Astra's robotics relevance extends beyond direct control. Models such as Astra are beginning to automate part of the work required to build the systems that move robots in the first place.
If that continues, the dedicated robot foundation model survives, but its place in the stack changes. The layer investors have treated as the strategic intelligence of the machine starts to look more like firmware: specialised, fast, embodied software that is generated, tuned and continually improved by a much larger system above it. That would move the centre of gravity of the robotics industry away from the model weights and towards the machinery that produces reliable behaviour.
Much of robotics happens away from the robot
OpenAI did not launch Astra as a robotics model. Its announcement focused on computer use, software engineering, browsing and professional work. Astra can install and test software, troubleshoot what it sees on screen and execute multistep computer workflows. It also posted a 95.9% geometric-overlap score on OpenAI's BenchCAD evaluation for reconstructing 3D objects by generating CAD code. (openai.com)
Robotics research is only intermittently robotic. Between physical trials, researchers read papers, edit Python, repair ROS stacks, inspect video, change reward functions, debug coordinate frames, launch reinforcement-learning runs, generate simulation assets, tune hyperparameters, compare checkpoints, analyse failures and decide what experiment to run next.
Increasingly, an agent can do that work.
ENPIRE, developed by researchers from NVIDIA, Carnegie Mellon and UC Berkeley, offers the clearest demonstration so far. The project starts from a familiar robotics problem: progress in real-world robot learning is limited by the amount of human supervision and algorithmic engineering each experiment requires. ENPIRE gives an AI agent access to the laboratory loop. It can reset the environment, run the robot, verify the result, inspect the evidence, alter the policy and try again. Its agents generate and revise policy code from rewards, video, traces and failure cases; consult research literature; modify training infrastructure; and compare competing approaches on physical hardware. (research.nvidia.com)
ENPIRE did not use Astra. The researchers evaluated Codex with GPT-5.5, Claude Code with Opus 4.7 and Kimi Code with Kimi K2.6. (research.nvidia.com) The underlying mechanism was already working before Astra arrived.
What ENPIRE has built is, in effect, an interface between an AI agent and physical evidence. Once a laboratory can reset a scene, observe an outcome, score it and expose those results safely to software, improvements in general-purpose coding and reasoning models can feed directly into robotics R&D. A human roboticist no longer has to mediate every improvement in general AI before it can be tested on a machine.
The robot model may become a component
Much of the venture case for robot foundation models rests on the idea that the model itself will remain the scarce and valuable part of the robotics stack.
Skild AI, for example, raised $1.4 billion in January at a valuation above $14 billion. Its stated ambition is a general-purpose robotic brain that transfers across quadrupeds, humanoids, arms and mobile manipulators. (skild.ai)
That bet has a serious technical basis. Physical action data is scarce. Contact dynamics are difficult. Robots operate in a distribution of friction, latency, geometry and failure modes that no corpus of web text contains.
Scarce data and difficult contact dynamics do not, by themselves, determine which layer captures most of the value.
Consider a stack in which a frontier generalist handles semantic reasoning, task decomposition, software engineering, experiment design, failure analysis and much of policy optimisation. The specialised embodied model below it then has a narrower job: translate intent into low-latency physical action reliably enough to satisfy the system above it.
Under that architecture, a robot model can improve technically while becoming less scarce strategically. A dedicated robotics company also faces an awkward competitive choice. It is unlikely to win a horizontal scaling contest against organisations training frontier models across vastly broader pools of compute, data, tools and economically valuable tasks. Moving down the stack towards specialised physical execution protects technical relevance, but it also makes the model look more like a component.
That makes the system that can generate, evaluate, deploy and improve those components a plausible source of bargaining power.
Specialised embodied pretraining works. OpenWAM-alpha, released this month, was pretrained on roughly 6,400 hours of robot and egocentric human data. On the real-robot bimanual portion of RoboDojo, it reported a score of 37.6 and a 24.4% success rate, versus 22.9 and 12.8% for the next-best model in its comparison. (openwam-official.github.io)
OpenWAM undercuts any claim that generic intelligence has already made robot-specific learning obsolete. It leaves open the question that matters for industry structure: whether technical necessity translates into strategic control of the stack.
If an agent above the action model can select architectures, build data pipelines, run ablations, modify rewards, inspect thousands of failures, choose training regimes, distil policies and deploy the winning candidate, then the policy becomes one output of a larger production system. A different body can get a different policy. A new warehouse task can trigger another training loop. A recurring failure can go back through the research system and produce another version.
The durable asset shifts from a single policy to the machinery that can repeatedly produce and verify good ones.
Physics still sets the boundary
Astra's own robot-arm results show why specialised execution is not about to vanish. It put a block in a bowl in 95% of trials, then fell to 10% when asked to seat a puzzle piece precisely into a groove. (openai.robocurve.org)
Precision exposes the part of robotics where language-model inference is weakest: high-frequency control under latency constraints, noisy sensing, continuous feedback and contact dynamics where a small positional error changes what happens next. An experimental study on a real UR5 robot found that promising offline VLA results could become unstable in closed-loop deployment because of coordinate conventions, temporal alignment, preprocessing and dataset coverage. Those are system-level problems, not simply a question of how capable the model is. (arxiv.org)
Leading robotics groups already build around this division of labour. Google DeepMind's Gemini Robotics ER 2 is a high-level embodied reasoning system. It interprets the world, plans multistep tasks and then hands motor execution to a lower-level Vision-Language-Action model. (blog.google) Skild uses a hierarchy too: a low-frequency high-level policy feeds a high-frequency lower-level policy that produces joint angles and motor torques. (skild.ai)
The likely architecture is hierarchical. Frontier models can sit above the reflex loop, where latency is less punishing and broad reasoning is more useful. They can design training curricula, study failures, rewrite reward functions and decide what the lower-level controller should learn next. The specialised model still executes the movement at the frequencies the physical system demands.
The frontier model is best placed in the coaching role, above the reflex loop, while the specialised controller stays on the field. Once the coach can redesign the training programme, the value of owning one particular controller changes.
The moat moves into the laboratory
If model intelligence becomes easier to obtain, the scarce parts of robotics are the things that remain stubbornly physical.
Robot-hours cannot be copied through an API. A calibrated manipulation station has to exist somewhere. Failed experiments consume time. Components wear out. Environments need resetting. Sensors drift. Safety boundaries need enforcement. Customer sites contain edge cases that never appear in a benchmark.
ENPIRE's laboratory infrastructure points to the same conclusion. Automatic reset, verification, logging, reproducible trials, parallel robot fleets, safe interfaces and auditable rollouts are what make physical experimentation usable by software at scale. (research.nvidia.com)
Software agents can break a build and revert the commit. In robotics, the world keeps the diff. The cup stays broken, the misplaced component remains misplaced, and a collision has already happened.
Physical AI therefore needs a hard boundary between probabilistic intelligence and deterministic authority. Collision envelopes, torque limits, watchdogs, permissions, verification routines and emergency interrupts have to remain outside the model's discretion. Safety and evaluation belong in the core product alongside the model.
For investors, parameter count and static benchmark leadership may become less informative than a simpler operational measure: time to trust. How many human hours and robot-hours does it take to move from a new task description to repeatable physical execution under real variation? How many manual resets are needed? How quickly does the system improve after failure? Can the learning transfer to another body or another site?
A company that compresses that cycle from months to days may own a stronger advantage than one that tops a single frozen manipulation benchmark.
The strongest case against this view is that embodied data itself could become the moat. A company running thousands of robots can accumulate proprietary action data that frontier-model providers do not have. High-frequency policies may also require architectures, latency profiles and hardware knowledge that remain deeply specialised. Robot-model companies can put frontier reasoning agents above their own control systems, and Google and Skild are already building hierarchical stacks.
If proprietary deployment data allows specialised models to widen their physical-performance lead faster than general agents can automate policy creation, the argument here weakens materially.
But even that outcome demands more than a model. The winning company would need the fleet, customer workflow, physical data engine, evaluation system, deployment infrastructure and feedback loop that let its model improve faster than competitors can reproduce it. In practice, it would have become a systems company.
Businesses whose imagined moat is simply that their foundation model knows robotics better are exposed to this shift. Its competitive threat comes from producing and improving the specialised systems beneath it, even if it never masters every motor action.
When AI automates the lab
The same logic changes the economics of the laboratory itself.
Software agents can be duplicated at software economics; physical laboratories cannot. As models improve at coding, visual reasoning, computer use, CAD and research, the cost of generating hypotheses and engineering experiments can fall much faster than the cost of running those experiments in the physical world. (openai.com)
As researcher attention becomes cheaper, robot throughput becomes the constraint. One engineer can supervise more experiments. More experiments can run in parallel. Successful recipes can spread across a fleet. The scarce resource becomes the physical infrastructure that can supply clean, safe, machine-readable evidence about whether a policy worked.
Generative AI has already made software cheaper to produce. In robotics, it may now make robotics research cheaper to produce as well.
Humans still specify objectives, constraints, safety requirements and accountability. Frontier agents can take over more of the digital research around the machine, while specialised models handle the physics. Over time, more of the work between those layers can become machine-to-machine.
For now, humans still occupy the decisive parts of the loop, and robot action models remain necessary. But if more experiment design, debugging, policy search and iteration can be delegated to agents, human attention moves towards choosing objectives, defining constraints and taking responsibility for systems that increasingly improve themselves.
The competitive metric in embodied AI may shift towards how quickly a company can turn general intelligence into reliable physical behaviour, measured in human and robot hours.
The companies with the strongest position will be the ones that control that loop: the physical data, the evaluation machinery, the safety system and the laboratory capacity in which robotic policies can learn quickly without being allowed to fail expensively.