Physical AI: How Machines Learned to See and Act — and Whether Robotics Has Had Its ChatGPT Moment

A useful definition of Physical AI is not “AI inside a robot.” It is an intelligent system closed around a physical feedback loop: it observes the world through sensors, estimates what is happening, chooses an action, changes the world through a body, and must then live with the resulting state. The body might be a six-axis arm, a wheeled warehouse platform, a car, a drone, a surgical instrument, or a humanoid. What makes the system physical is the loop, not the silhouette.

This separates Physical AI from two neighboring categories. Fixed automation can be extremely sophisticated, but it performs a known sequence in an environment engineered to remove uncertainty: a fixture locates the part, a light curtain protects the cell, and a controller repeats a validated trajectory. A digital agent can work in a broad information space, but its actions usually modify text, code, tickets, or API state. Those changes can be copied, reviewed, rolled back, or replayed at low cost. A physical policy operates on a world that is only partly observed, continuously changing, and expensive to reset.

The categories overlap in a real product. A robotaxi contains learned perception and prediction, route planning, classical control, maps, remote assistance, and deterministic safety mechanisms. An industrial arm may use a learned vision model to locate randomly piled parts and a conventional motion planner to pick them. “Physical AI” therefore names a systems property, not a claim that every component is a neural network.

It also helps resolve an argument about whether autonomous cars belong in robotics. They do. A car senses, predicts, acts through a body, and carries physical risk. Robotaxis are especially important because they show that the field can reach a ChatGPT-like adoption threshold vertically: an ordinary customer requests a ride without learning robotics, while an enormous operational system hides the complexity. That is a stronger test than whether a laboratory demo understands an impressive spoken command.

The field did not suddenly begin when a language model was connected to a robot. Its central question has been restated for seventy years: how much structure should an engineer put into the machine, and how much should the machine acquire from experience? The history is easier to read as four overlapping acts than as a procession of humanoid prototypes.

Act I, 1948–1973: program the physical world

Norbert Wiener’s 1948 Cybernetics made feedback a common language for machines, animals, and organizations: do not merely issue a command; measure the consequence and correct the next command. Grey Walter’s small “tortoises” showed how simple sensor loops could produce apparently purposeful movement. These were not intelligent in today’s sense, but they established the primitive from which Physical AI still cannot escape: action and observation are coupled.

George Devol filed the patent behind Unimate in 1954, and the first production unit went to General Motors in 1961. The industrial robot was transformative precisely because the world around it was made predictable. It did not need to understand a factory. Engineers constrained the cell until a programmable mechanism could repeat a valuable motion. That bargain—high capital and integration cost in exchange for high repeatability— still describes much of the industrial-robot market.

Shakey, built at SRI between 1966 and 1972, attempted the opposite bargain. A mobile robot combined cameras and range sensors with a world model, symbolic goals, route search, and actions. Its work helped produce techniques such as A* search and STRIPS planning. Shakey could reason about a prepared room, but perception was slow, the environment sparse, and the symbolic model brittle. It established the classical “sense–model–plan–act” architecture and exposed its cost: an explicit model has to stay aligned with a disorderly world (SRI, Shakey the Robot).

Waseda’s WABOT-1, completed in 1973, integrated locomotion, manipulation, vision, and language-like interaction in a human form. Its importance was architectural rather than commercial: the humanoid body was already understood as a systems-integration challenge. Fifty years later, the body has better actuators and learned policies, but balance, energy, manipulation, perception, and safety still have to work simultaneously.

Act II, 1986–2007: act under uncertainty

Rodney Brooks argued that intelligence need not begin with a complete internal reconstruction of the world. His subsumption architecture layered fast behaviors—avoid, wander, explore—so a robot could respond before a central planner finished reasoning (Brooks, 1986). Reactive robotics traded global elegance for timely competence. Modern learned systems have not eliminated this insight: a robot still needs fast loops beneath slow semantic reasoning.

In the late 1980s, Carnegie Mellon’s ALVINN learned steering from road images with a small neural network. In 1995, the Navlab project drove “No Hands Across America,” with neural steering for most of the route while humans retained throttle and braking. The demonstration was narrow and heavily supervised, but it introduced a durable pattern: learn a sensor-to-action mapping from human examples instead of writing every visual rule.

The DARPA Grand Challenges then forced autonomous driving out of controlled demonstrations. In 2004, no entrant finished the desert course. In 2005, five did; Stanford’s Stanley completed 132 miles in 6 hours and 53 minutes. The 2007 Urban Challenge added traffic rules and interactions. Progress came not from one universal neural model but from better sensors, probabilistic localization, perception, planning, control, maps, compute, and relentless field testing (DARPA’s program history). This systems lesson would later be obscured by the fluency of end-to-end demos.

Act III, 2012–2020: replace hand-built features with experience

AlexNet’s 2012 ImageNet result did not solve robotics, but it changed its most information-starved component: perception. Objects, lanes, people, poses, and depth could increasingly be learned from data rather than described by handcrafted features. Cheap cameras, GPUs, large labeled datasets, and convolutional networks shifted the bottleneck from “can the machine see this at all?” toward “can it act robustly on what it sees?” (Krizhevsky, Sutskever, and Hinton, 2012).

Robot-learning teams built data factories. Google researchers used multiple arms to collect hundreds of thousands of grasp attempts, then let experience from one robot improve the others. QT-Opt later combined off-policy reinforcement learning with 580,000 real-world grasp attempts and reported high success in a deliberately bounded picking setup. The exact benchmark matters less than the operational pattern: a fleet could turn action into a reusable training corpus (Google Research, QT-Opt).

Simulation added cheap failure and reset. Domain randomization varied textures, lighting, geometry, and dynamics so a policy could no longer depend on one synthetic world. OpenAI’s Dactyl work showed an anthropomorphic hand transferring difficult manipulation policies from randomized simulation to hardware (OpenAI, 2018). Yet sim-to-real remained a negotiation with the reality gap, not a magic export button.

Autonomous-driving programs used the same decade to turn research stacks into fleets. Waymo began offering fully driverless rides to the public in Phoenix in 2020. That milestone already contained the template for mature Physical AI: constrain the operating domain, instrument the fleet, build remote and local fallback, validate safety, and expand geography only when the whole system is ready.

Act IV, 2022 onward: train the task interface and the action policy together

SayCan connected a language model’s knowledge about plausible plans to the affordances of robot skills. Gato represented text, images, controls, and actions in one sequence model. RT-1 then trained a transformer policy across hundreds of real robot tasks. RT-2 co-trained web-scale vision-language data with robot action data. Open X-Embodiment pooled trajectories from many laboratories and bodies. The unit of engineering began to change from a task-specific pipeline to a model, dataset, and post-training system that could support a family of tasks.

This is the genuine historical break. It does not mean symbolic planning, motion planning, calibration, control, maps, or safety disappeared. It means semantics and action could share a learned representation, and knowledge acquired outside one robot task could improve behavior inside it. The next two sections explain why that is a foundation-model moment—and why it is not yet the same event that ChatGPT was for general users.

A text model can consume trillions of tokens copied from an existing network. A robot trajectory has to be produced by a body in time. Someone buys the hardware, prepares the scene, supervises the attempt, resets the objects, labels success, repairs wear, and accepts that a failed action may break the very thing the model was meant to handle. Physical data are not just scarce; their acquisition is an operational process.

Data are expensive and causally incomplete

Video shows what an action looked like, but usually not the force, joint state, tactile contact, failed attempt, or alternative action needed for control. Teleoperation provides aligned observations and actions, but one hour of demonstration takes roughly one hour of operator time before setup and curation. Autonomous collection scales better only after the policy is safe enough to attempt the task and the system can recover or request help.

Bodies are not interchangeable tokenizers

Text models share a relatively stable interface: tokens in, tokens out. Robots have different cameras, joint limits, grippers, control rates, payloads, kinematics, and failure modes. The same instruction—“put the cup on the shelf”—maps to different reachable spaces and action coordinates on a tabletop arm, a mobile manipulator, and a humanoid. Cross-embodiment training can learn common semantics and some reusable structure, but embodiment metadata and hardware-specific post-training do not disappear.

Real time is part of correctness

A late braking decision is a wrong decision. A balance controller cannot wait for a large model to deliberate. A gripper may need to react to slip faster than a semantic policy can process another camera frame. This creates a multirate architecture: slow reasoning, medium-rate action generation, fast control, and continuously enforced safety. Latency, jitter, thermal limits, network loss, and edge compute are therefore model-quality concerns, not merely deployment details.

Contact is discontinuous and only partly observed

Language and pixels describe smooth statistical regularities. Contact introduces sharp transitions: a plate sticks, a cable snags, a drawer jams, a tire loses grip, an object shifts behind the hand. Small state-estimation errors can produce very different outcomes. Tactile sensors help, as do force control and learned recovery, but the training distribution rarely contains every material, geometry, wear state, and human interaction.

Errors can be irreversible

A hallucinated paragraph is inconvenient; a hallucinated free path can injure someone. Physical systems therefore require an evidence hierarchy that digital demos can postpone: hazard analysis, speed and force limits, protected stops, redundancy, monitored degradation, incident review, and exposure-normalized statistics. Safety also changes the training loop because the most informative edge cases are often the ones a real system must not explore freely.

Long tasks multiply small weaknesses

A policy can look excellent on a single pick and still be useless for a twenty-minute job. As a toy illustration—not a model of a particular robot—if each of 100 dependent steps succeeds with 99% probability, the whole chain succeeds only about 37% of the time: 0.99¹⁰⁰ ≈ 0.37. Real errors are correlated and recovery can improve the result, but the arithmetic explains why long demonstrations, intervention rates, and self-recovery matter more than a montage of clean subtasks.

These frictions do not imply that general robotics is impossible. They explain why its scaling curve is slower and more operational than language modeling. The decisive innovation must improve not only the policy, but also data collection, control, safety, maintenance, and the economics of each useful action.

RT-1 is the cleanest starting point. Google trained a transformer on roughly 130,000 episodes gathered over seventeen months by a fleet of thirteen Everyday Robots platforms, covering more than 700 tasks. Images and a natural-language instruction entered the model; discrete action tokens came out. The important result was not that a transformer could move one arm. It was that one policy could absorb a broad, uneven task corpus and improve as task and data diversity grew (Google Research, RT-1).

RT-2 changed the pretraining equation. It started from a vision-language model exposed to web data, represented robot actions as another output vocabulary, and co-trained vision, language, and robotic trajectories. The web corpus could provide semantic concepts that a laboratory dataset was unlikely to contain in sufficient variety. The robot could use those concepts to interpret novel instructions and combinations, then express a grounded response as action tokens (Google DeepMind, RT-2).

This is easy to misstate as “the internet taught the robot to move.” It did not. Web pretraining supplied visual and linguistic abstractions; robot trajectories supplied executable motor grounding. RT-2 could recombine known skills under new semantic conditions, but it could not manufacture a dexterous motor primitive that was absent from the action data. The distinction is fundamental: semantic breadth and physical competence scale through different sources, joined inside the policy.

Open X-Embodiment attacked the second shortage: no single laboratory had enough diverse robotic experience. The collaboration aggregated more than one million trajectories from 22 embodiments, 527 skills, and 160,000 task formulations, then trained RT-X models across that mixture. Cross-embodiment training improved transfer in the participating evaluations, suggesting that heterogeneous trajectories can form something like a shared robotics corpus (Open X-Embodiment collaboration).

“Something like” matters. Common Crawl stores text through a broadly shared interface. Open X has to reconcile camera placements, action spaces, control rates, grippers, task labels, and data quality. A trajectory from one body is not automatically executable on another. What transfers most readily is the latent structure of objects, goals, spatial relations, and common manipulation patterns; embodiment-specific control still requires adapters, normalization, or post-training.

Three design changes nevertheless survived. First, a natural instruction could select and compose behavior rather than merely trigger a hard-coded skill name. Second, action could be treated as a learned output modality, alongside text and image representations. Third, a policy could be pretrained once and adapted across a family of tasks, making model quality, dataset diversity, and training infrastructure shared platform concerns.

“ChatGPT moment” is often used to mean “a demo that surprised people.” That bar is too low. ChatGPT combined a capability threshold with a natural interface, immediate access, a feedback loop, and usable economics. It let millions of non-specialists test a general model on their own work. A fair robotics analogy must therefore ask about a product and operating system, not only a model paper.

1. Generalization: real, but bounded by physics and bodies

Modern policies can follow unseen linguistic formulations, recognize new objects, and combine known skills in new scenes. Models trained across robots can outperform single-embodiment baselines. Yet a change in lighting, gripper, contact geometry, payload, clutter, or task horizon can still move the system outside its competence. Generalization is no longer absent; it is uneven and difficult to specify.

2. Natural interface: substantially crossed

Language and demonstration have become practical task interfaces. A user can describe a goal rather than choose a numbered skill, while a planner can convert the instruction into subgoals and a VLA policy can ground them. The caveat is that a fluent response can disguise a narrow action repertoire. Understanding “clean the kitchen” is not the same as possessing every manipulation, navigation, recovery, and safety behavior that the instruction implies.

3. Data flywheel: emerging, not frictionless

Fleet logs, interventions, teleoperation, simulation, and human video now feed shared models. AutoRT used more than fifty robots to collect 77,000 episodes in varied office settings. DROID pooled 76,000 trajectories, 350 hours of interaction, and 564 scenes across institutions (DROID dataset). These are meaningful corpora, but they remain tiny and operationally expensive beside web text. More deployment does not automatically create clean labels or safe exploration.

4. External adoption: crossed in vertical systems

Robotaxis pass a powerful version of this test. A rider requests transport through an ordinary app; no robotic expertise is required. Waymo reported more than 400,000 rides per week in February 2026. Baidu reported an Apollo Go weekly peak above 350,000 fully driverless rides in March. Both are company-reported operating metrics, and neither implies autonomy on every road. They do show real customers repeatedly buying a physical service.

Industrial and consumer robotics crossed narrower thresholds even earlier. Factories have deployed large robot populations for decades; floor-cleaning robots are mass consumer products. What changed in robotaxis is the combination of learned autonomy, public access, and fleet scale inside an operating domain that is broad enough to feel like a service rather than a machine cell.

5. Reliability and economics: the horizontal barrier

A general robot must perform useful work for long periods, recover from predictable failures, expose a safe fallback, and cost less than the process it replaces or enables. Today’s public humanoid evidence is dominated by vendor-selected tasks and videos. Factory pilots are valuable, but a successful insertion or tote movement is not yet a multi-customer reliability distribution. Homes add unstructured objects, children, pets, privacy, remote assistance, and a much weaker service infrastructure.

The resulting verdict has three levels. RT-2 was the GPT-3 moment for robot policy architecture. Open X-Embodiment was an early ImageNet-like moment for a shared training corpus. Waymo and Apollo Go have had a vertical ChatGPT moment inside selected geographies. General-purpose robotics has not had the horizontal product moment yet. It is building the platform on which one could occur.

The popular mental image is a multimodal LLM watching a camera and emitting motor torques. That is not a responsible description of a production architecture. Semantic reasoning and precise control have different data, latency, and safety requirements. The modern pattern separates them while allowing learned representations to cross the boundaries.

The slow layer answers “what should happen next?”

A vision-language model or embodied-reasoning model interprets the instruction, scene, history, and constraints. It may decompose “clear the table” into object-level subgoals, query a map, select a tool, or ask for clarification. Memory belongs here because the system must know which subgoals have been attempted and why a previous route failed. This layer can use substantial compute because its output changes at the pace of a task, not a joint.

The action expert answers “how should this short segment move?”

A VLA policy receives images, proprioception, and a goal, then produces a short horizon of actions. Diffusion policies and action-chunking transformers are attractive because manipulation is multimodal: there may be several valid trajectories, and committing one small increment at a time can amplify noise. Predicting a chunk provides temporal consistency; receding-horizon execution lets the system observe again before consuming the entire plan.

The action representation varies. A tabletop policy may output end-effector deltas and gripper commands. A mobile robot may produce waypoints. A whole-body humanoid policy may emit target poses or latent controls. The more directly a model controls the body, the more its training data, inference latency, and safety envelope must reflect that body.

The fast layer keeps the body physically coherent

Classical control, model-predictive control, and learned low-level policies track the target, maintain balance, limit force, and react to perturbations. Figure describes Helix 02 as a three-level system: a semantic system, a full-body policy running around 200 Hz, and a learned controller around 1 kHz. Those rates and the demonstrated behavior are vendor claims, but the decomposition illustrates why a large semantic model cannot be the only controller (Figure, Helix 02).

The safety layer has authority over every learned layer

A production system needs constraints that do not depend on the model agreeing with them: speed and force caps, collision zones, monitored stops, brake and steering limits, state-of-health checks, geofences, redundancy where the hazard requires it, and a defined degraded mode. An independent supervisor should be able to reject or reshape a plausible but unsafe action. The exact mechanisms differ by machine, but “the model was confident” is never a safety case.

Gemini Robotics makes the division visible from another angle. Google DeepMind introduced a VLA model for direct action and a separate embodied-reasoning model for spatial reasoning, planning, and interaction with conventional controllers. Gemini Robotics 2, announced on July 30, 2026, pushes further toward integrated whole-body behavior. The published task results are internal and vary widely by task, so they are evidence of breadth, not a fleet service-level objective (Google DeepMind, Gemini Robotics 2).

The architectural consequence is practical. Teams no longer have to choose between an entirely handcrafted stack and an opaque end-to-end policy. They can learn the layers where data create an advantage, retain explicit control where physics and assurance demand it, and define contracts between the rates. The research frontier is not “remove all structure.” It is “learn the right structure without losing observability and authority.”

Foundation models shift the central asset from a library of task programs to a data-and-evaluation loop. The loop is more diverse than “collect demonstrations, imitate them.” Each source supplies a different missing variable; no single source is sufficient.

Internet video buys semantics, not executable contact

Human video supplies enormous visual diversity: kitchens, tools, clothing, hands, clutter, and goal-directed sequences. It can help a model infer that a mug is grasped by the handle or that a drawer must open before an object is placed inside. But ordinary video does not reveal joint angles, camera calibration, fingertip force, or the counterfactual action that would have prevented a slip. Video pretraining is therefore a semantic prior and a source of behavior structure, not a replacement for robot data.

Demonstrations establish grounding

Teleoperation records the observation, robot state, and action together. Better interfaces—VR controllers, exoskeletons, leader–follower arms, instrumented grippers, or direct human-video transfer—can increase quality and reduce embodiment mismatch. Demonstrations are still biased toward what an operator can do smoothly. They often underrepresent hesitation, recovery, ambiguous instructions, and the policy’s own future mistakes.

Dataset collaboration spreads the fixed cost. DROID gathered data with a common portable setup across institutions. Open X-Embodiment normalized many existing datasets. AgiBot World reports more than one million trajectories across 217 tasks; that scale is a dataset-provider claim and should be evaluated through coverage, sensors, labels, and downstream transfer rather than trajectory count alone (AgiBot World).

Simulation manufactures coverage and safe failure

A simulator can run thousands of environments in parallel, reset instantly, expose privileged state, and sample rare or dangerous conditions. Reinforcement learning can use that scale to acquire locomotion and control behaviors impractical to learn through unrestricted hardware exploration. Digital twins can reproduce a warehouse or road geometry before deployment.

The constraint is epistemic: a policy cannot learn the part of physics the simulator gets wrong. Domain randomization reduces dependence on one rendering or parameter set. System identification aligns selected dynamics. Real-world fine-tuning corrects residual errors. The robust pattern is not “simulation instead of reality,” but an explicit ladder from broad synthetic coverage to progressively more expensive real evidence.

World models add learned counterfactuals, not ground truth

A world model predicts future observations or latent states under candidate actions. It can support planning, generate visual variation, compress experience, and expose a policy to scenes that were not recorded exactly. NVIDIA’s Cosmos platform explicitly treats generative world models as part of the Physical AI data stack; GR00T connects that synthetic-data approach to humanoid policies (NVIDIA Cosmos documentation).

A visually plausible rollout can still contain wrong friction, mass, collision, or causal response. World models are useful for proposal generation, representation learning, and stress-test coverage. They are not safety evidence by themselves. Contact-rich tasks should combine learned worlds with physical simulation, domain randomization, hardware tests, and calibrated uncertainty.

Autonomous attempts reveal the policy’s own distribution

Once deployed, the system encounters states that a demonstrator did not choose. Human takeovers, remote assists, failed grasps, protected stops, and recovery sequences are disproportionately valuable because they map the boundary of current competence. DAgger-style data aggregation, offline reinforcement learning, and carefully constrained online learning all try to exploit this distribution without letting unsafe exploration dominate.

The flywheel closes only when operations are instrumented. A fleet must record the instruction, relevant observations, model and software version, action, intervention, outcome, environment, latency, and maintenance state. Otherwise deployment produces a video archive rather than a learning corpus. The deepest moat may therefore be an operating discipline: the ability to turn every useful hour and every interruption into comparable evidence.

A humanoid is not automatically the most efficient machine. It is a wager on compatibility. Human buildings already contain stairs, doors, shelves, tools, carts, workstations, and aisle widths designed for our reach and proportions. If a robot can use that infrastructure, a customer may avoid redesigning the site around a fixed cell or a specialized conveyor.

The body is also a data interface. Human motion, egocentric video, and teleoperation map more naturally to a roughly human morphology than to a bespoke mechanism. A single platform could in principle move between material handling, machine tending, inspection, and simple household work, allowing one model and fleet system to amortize integration across tasks. That is the economic thesis behind the current investment wave.

Every advantage has a physical price. Bipedal balance consumes control authority and energy. A fall can damage the robot or the environment. Hands add many degrees of freedom and fragile contacts. Head-mounted cameras move with the body. Batteries must cover a useful shift while actuators manage heat. A machine sharing human space needs conservative force and speed limits, which can erase the throughput advantage promised by a fast demo. Maintenance becomes a distributed mechanical service problem.

This is why the right comparison is not “humanoid versus human.” It is “humanoid versus the cheapest safe process redesign.” A mobile base with one arm may handle tote movement more reliably. A fixed cobot may deliver higher precision. A conveyor may remove the manipulation problem entirely. Morphology is an economic and systems decision, not a philosophical destination.

Google DeepMind’s Gemini Robotics program is important because it combines broad multimodal priors with embodied reasoning and direct action, and because it evaluates across multiple robot forms rather than equating Physical AI with humanoids. Its March 2025 launch demonstrated adaptation to new objects and tasks; Gemini Robotics 2 expanded the architecture in July 2026. The reported success rates are internal and differ sharply by task, so the result is an early capability frontier, not a service-level benchmark (Google DeepMind, Gemini Robotics).

NVIDIA’s GR00T N1 is an open foundation model for humanoid reasoning and skills, paired with Isaac simulation, synthetic motion generation, and Cosmos world models. This is valuable less as a declaration that humanoids are solved than as an attempt to standardize a development platform. NVIDIA’s own documentation makes the current boundary visible: adaptation still assumes task demonstrations; persistent memory, arbitrary zero-shot embodiment transfer, and autonomous self-retry are not general solved capabilities (NVIDIA Research, GR00T N1).

Physical Intelligence’s π0 combined an approximately three-billion- parameter vision-language backbone with a roughly 300-million- parameter continuous-action expert. Later work emphasized open-world generalization, cross-embodiment post-training, hierarchical reasoning, memory, and reinforcement learning. The architectural line is credible: a shared semantic backbone plus an action model can be adapted to several bodies. The public evidence is still primarily research evaluation and selected company demonstrations, not an independently measured labor service (Physical Intelligence, π0).

Figure’s Helix 02 demonstrates why humanoid progress now looks qualitatively different from a sequence of scripted clips. The company showed a roughly four-minute, 61-action kitchen task with full-body control and presented a multirate architecture trained with human motion and large simulated experience. This is a vendor demonstration: no independent benchmark establishes the intervention rate, repeated-run distribution, damage rate, energy cost, or maintainability. It is strong evidence that longer integrated behavior is becoming possible, not that a household product is finished.

The most honest reading of the humanoid wave is therefore “early platform proof.” The model families show that semantic transfer, whole-body policies, action chunks, simulation, and human-motion data can be integrated. The market proof will arrive only when customers publish sustained throughput, intervention, safety, service, and cost results across sites.

The attention market and the deployment market are different. Humanoids dominate video because viewers can immediately read the difficulty of a human-shaped action. The largest installed bases still belong to specialized machines whose task, body, and environment were designed together.

The International Federation of Robotics counted 542,000 new industrial-robot installations in 2024, the second-highest annual total on record. China accounted for 295,000 installations, or 54% of the world total, and domestic Chinese suppliers reached a 57% share of their home market. These are industry-association statistics, not a census of foundation-model robots. They demonstrate something more basic: physical automation is already a huge capital system, and new AI must integrate with it (IFR World Robotics 2025).

IFR’s service-robot chapter reports almost 20 million consumer service robots sold in 2024, overwhelmingly floor-cleaning systems, and 102,900 professional transport and logistics robots. The professional figure comes from a sample of suppliers and is not the complete market. That limitation is useful: robotics numbers often combine units sold, robots installed, subscriptions, rides, miles, and vendor samples. A serious market comparison keeps the unit and denominator attached.

Factories and warehouses: intelligence is a workflow property

An industrial arm succeeds because grasping, fixturing, process timing, quality control, and human access are jointly engineered. Learned vision expands the range of parts and placements; learned policies can reduce task-specific programming. But the highest value often comes from changing the surrounding process so the robot sees fewer ambiguous states. Foundation models do not repeal industrial engineering.

Amazon announced its millionth deployed robot in June 2025 across more than 300 facilities. It also claimed that DeepFleet, a model coordinating robot movement, improved fleet travel time by 10%. Both are vendor-reported metrics. The case is important because the model’s body is the fleet: routing and orchestration can create more value than a more anthropomorphic unit (Amazon, 2025).

Robotaxis: general software inside a constrained domain

Waymo’s February 2026 funding announcement reported more than 400,000 paid rides a week, 15 million rides during 2025, more than 20 million lifetime rides, and 127 million fully autonomous miles. The company also reported a 90% reduction in serious-injury crashes relative to a human benchmark over that exposure. The rides and miles are vendor operating metrics; the safety comparison has a disclosed methodology and denominators but remains a company analysis rather than a universal regulator verdict (Waymo, 2026 funding and operating update).

Baidu’s Q1 2026 results reported 3.2 million fully driverless Apollo Go rides in the quarter, a weekly peak above 350,000 in March, more than 22 million cumulative rides by April, and a footprint spanning 27 cities in May. These are unaudited company operating metrics in an investor release. They nonetheless show that the vertical threshold is not unique to one US fleet (Baidu, Q1 2026 results).

A robotaxi is broad across scenes and narrow across purpose. It handles an enormous variety of roads, actors, and weather inside a declared operational design domain, but its objective, body, regulations, maps, service process, and fallback are fixed. That is exactly why it can be mature. Restricting the domain is not a failure of intelligence; it is the systems mechanism that turns a probabilistic model into an accountable service.

Driver-assistance mileage under active human supervision should not be merged with rider-only robotaxi mileage. The former may be a valuable product and a source of data, but the human retains the fallback role. A useful autonomy taxonomy must state who is responsible for monitoring, when remote assistance occurs, and whether the vehicle can complete the trip without a safety driver.

Who is missing from this map

This survey has a deliberate boundary: it relies on actors whose work can be read through public, checkable evidence—primary papers, IFR statistics, vendor claims with a stated methodology. Several major names do not fit that frame, and passing over them silently would turn a sampling constraint into a hidden verdict.

First, Tesla. The company operates what is probably the largest supervised-autonomy fleet in existence and is developing the Optimus humanoid. Its public numbers, however, are built around supervised miles—and, as the previous paragraph argues, mileage with a human in the fallback role and rider-only mileage are different measured quantities. Until disclosure supports a “driverless rides and interventions per unit of exposure” row, Tesla has no line in the verdict tables. That is a limit of the method, not a judgment of the technology.

Second, the Chinese humanoid wave: Unitree, UBTech, AgiBot, and others already ship production hardware rather than only videos. This article cites AgiBot solely as a dataset provider, yet the company is first of all a robot manufacturer. Structurally, the wave rests on a fact the article has already recorded: more than half of global industrial-robot installations happen in China. The manufacturing base exists; what is missing is public operating evidence at the level of the evidence ladder used here.

Third, Boston Dynamics, which set the bar for dynamic whole-body control for two decades: Atlas and Spot showed what is physically possible long before the current wave. The company rarely publishes interventions-per-hour-style metrics, so it also has no row in the verdict rubric. That absence is a reminder about the rubric itself: the map in this article describes proven useful work, not the field’s full engineering achievement.

The foundation-model thesis changes the venture story. A company is no longer presented as selling one robot for one task; it is presented as building a reusable “brain,” a data engine, a manufacturing system, and a fleet platform. If that platform generalizes, the addressable market expands from a machine category to a large share of physical labor and mobility. That possibility can rationally attract very large capital before the product is mature. It cannot substitute for evidence that the platform generalizes.

Figure announced more than $1 billion in Series C commitments at a $39 billion post-money valuation in September 2025 (company announcement). Skild AI announced a $1.4 billion Series C at a valuation above $14 billion in January 2026 (company announcement). Apptronik said its February 2026 extension brought total Series A funding above $935 million (company announcement). These numbers measure investor commitment and financing capacity. They do not measure successful operations per hour.

Waymo’s $16 billion round at a $126 billion valuation is different in one important respect: the same announcement included a large paid-ride base and autonomous-mile exposure. That still does not reveal unit economics, but it places the company higher on an evidence ladder than a startup whose public proof is selected pilots. Capital and deployment evidence can coexist; they should not be collapsed into one signal.

Forecast disagreement is information

Goldman Sachs estimated in 2024 that the humanoid-robot market could reach $38 billion by 2035 and 1.4 million annual unit shipments. Morgan Stanley estimated in 2025 that a humanoid market could reach $5 trillion and one billion robots by 2050. These are analyst forecasts, not observed market facts. They use different horizons, market boundaries, price paths, adoption curves, and assumptions about which kinds of labor count (Goldman Sachs, 2024; Morgan Stanley, 2025).

Their wide divergence should not be averaged into a more “reasonable” number. It is the result: the market definition and technical adoption path are unresolved. A useful forecast must state whether it counts hardware revenue, robot-as-a-service, software, maintenance, enabled labor value, or all of them; whether the unit is a shipment, installed base, or productive equivalent; and what autonomy, utilization, and price assumptions connect the technical curve to revenue.

An evidence ladder for buyers and technical leaders

At the bottom is capital: it buys time, compute, robots, and manufacturing capacity. Next is a selected demonstration: useful for revealing a possible behavior. Repeated internal tests add a distribution. A customer pilot adds a foreign environment and workflow. Multi-site operation tests transfer. Sustained paid work adds economic pressure. Independent safety and performance evidence makes the claim challengeable.

Each rung answers a different question. A four-minute kitchen demo asks, “Can this architecture produce a long sequence under these conditions?” A pilot asks, “Can it fit one workflow?” A fleet SLO asks, “How often does it work, fail, need help, and recover?” Unit economics asks, “Does the useful output justify hardware, supervision, energy, service, and downtime?” Investment headlines answer none of the last three.

Physical AI is not one market moving at one speed. Maturity rises when the task is valuable and repeatable, the environment can be constrained, failure is detectable, a human fallback exists, and the data loop is dense. It falls when the task is open-ended, contact-rich, safety-critical, geographically diverse, and weakly instrumented.

Industrial arms in guarded or collaborative cells are mature for defined operations. Warehouse transport and sorting are mature in engineered facilities. Consumer floor cleaning is mature as a narrow task. Robotaxis are commercially mature in selected operational design domains and immature outside them. Drones are technically strong in mapping, inspection, and constrained delivery, while regulation, weather, communication, and population exposure often bound deployment. Agricultural robots succeed where crop, terrain, season, and operation are specific; biological variation prevents one blanket rating.

Medical robotics is mature as surgeon-controlled instrumentation and assistance, but autonomous intervention has a much higher evidence threshold. General mobile manipulation in factories is entering paid pilots. General home assistance remains early because every object, person, room, and risk policy changes. The humanoid body appears in several of these domains, but it does not determine maturity by itself.

Start with the operational design domain

Before comparing models, write the domain in operational terms: locations, surfaces, lighting, weather, object families, payloads, people, task horizon, network conditions, acceptable speed and force, maintenance access, and fallback. A claim such as “95% success” is uninterpretable without this boundary and a sampling process. Expansion should be measured as a change to the domain, not narrated as a vague move toward generality.

Measure a process, not a highlight reel

The primary numerator is useful output: a completed ride, pick, inspection, assembly step, room-cleaning cycle, or delivered item that meets quality criteria. The denominator is exposure: hours, kilometers, operations attempted, sites, shifts, or interactions. Report both. A robot that completes 99 of 100 easy picks and none of the rare critical ones may be worse than a slower system with a safe, cheap fallback.

Intervention deserves its own taxonomy. A planned operator check is not an emergency takeover. A remote semantic hint is not a low-level teleoperation segment. A protected stop is not a failed completion, but it consumes time and may reveal a safety boundary. Hidden remote labor can make an autonomous product look more mature than it is; transparent assistance can make a service much safer. The distinction is disclosure, not whether assistance ever occurs.

Price the whole useful operation

Total cost includes the robot and compute, site integration, teleoperation, network, charging, batteries, spares, field service, software, insurance, safety process, and lost production during downtime. Divide that cost by accepted output, not calendar time. A humanoid that can perform many tasks may justify higher hardware cost if changeover is genuinely cheap; a specialized machine wins if its throughput and reliability dominate one stable bottleneck.

Test adaptation as a marginal cost

Foundation-model claims should reduce the cost of the next task. Record how many demonstrations, simulation hours, engineer-days, prompts, adapters, safety cases, and site changes were required to move from task A to task B. Then repeat on a new body or site. If each deployment is a bespoke robotics project, the shared model may still be useful, but it has not yet created platform economics.

Demand evidence that survives outside the vendor

The strongest practical sequence is staged. Reproduce the claimed task internally. Run repeated trials with prespecified success and intervention definitions. Add adversarial but in-domain variation. Operate on the customer process for complete shifts. Publish or contract against the resulting distribution. For safety-critical systems, add third-party, regulator, or standards-based evidence. A model card and a cinematic video are useful inputs to this process; they are not substitutes for it.

1. 01

   The breakthrough is architectural, not anthropomorphic

   Physical AI crossed a foundation-model threshold when language, vision, state, and action could be trained inside reusable policies, and when heterogeneous trajectories began to form a shared corpus. The important object is the learning and operating stack—not whether the body has two legs.

2. 02

   RT-2 was the GPT-3 moment; Open X began the corpus moment

   RT-2 made web semantics available to action and revealed a credible scaling direction. Open X-Embodiment showed that robot data from many institutions and bodies can improve shared policies. Those advances changed the research default, but they did not create a product that any person could deploy on any task.

3. 03

   Robotaxis have already crossed a vertical ChatGPT threshold

   Waymo and Apollo Go hide a deep autonomy stack behind an ordinary ride request and operate at meaningful public scale. Their success comes from a broad policy inside a tightly engineered purpose, geography, service process, and safety system. That is not lesser intelligence; it is how probabilistic capability becomes a product.

4. 04

   The durable design is hybrid and multirate

   Slow semantic planning, a learned action expert, fast body control, and an independent safety envelope solve different problems. LLM fluency is not torque control, and an end-to-end policy is not an excuse to remove authority, monitoring, or deterministic constraints. The best systems learn aggressively while preserving explicit contracts around physics and risk.

5. 05

   The remaining threshold is verified useful work

   General robotics will have its horizontal ChatGPT moment when a non-specialist can assign a materially new physical task and receive sustained, safe, economical output without a bespoke integration project. Until then, autonomous useful time, interventions, throughput, cost, energy, service, safety, adaptation, and independent verification are better signals than funding, model size, or demo fluency.