Where Reason Ends and Act Begins: A July 2026 Progress Report on Physical AI

For three years, "physical AI" mostly meant a video. A robot folding a shirt, a humanoid walking down stairs, a car navigating a closed course - impressive, carefully lit, and, more often than not, teleoperated or choreographed well outside the messiness of a real workplace. 2026 is the year that stopped being enough. Package counts, production throughput, IPO prospectuses, and state-mandated safety filings have replaced the demo reel as the unit of proof, and the number of physical AI systems now generating that kind of evidence - rather than another video - has gone from a handful to an entire sector in the space of about eighteen months.

This article is part of a series exploring the Physical AI and AIoT landscape, and it is a different kind of entry than the others. Where the earlier pieces built out the Sense - Reason - Act framework conceptually, this one is a progress report: a check-in on how far the Act pillar has actually moved in the real world, using the evidence that has accumulated over the past year rather than the theory. If you have not read the earlier pieces - From Code to Concrete: The Rise of Physical AI and the Power of Convergence, Data Residency, Data Sovereignty, and the Rise of Inference Sovereignty, Data Has Currency: The Case for the Internal Data Brokerage, and Beyond the Language Model: Why World Models Are the Next Frontier for Physical AI - I would encourage you to do so, because the Act pillar is the point where a decision leaves the realm of information and becomes physical consequence, and for the first time there is a genuine wave of commercial evidence - not just research papers - showing what that looks like at scale. Next month's companion piece turns to the accountability question that evidence raises. This one is about the evidence itself.

1. From Demo to Deployment

The clearest sign that something has shifted is not any single robot's capability - it is what companies are now willing to be measured against. A choreographed demo can be re-shot until it works. A production line cannot. Figure AI's humanoid has been running real shifts on an active BMW assembly line in Spartanburg, South Carolina since 2025, and by mid-2026 the company's BotQ manufacturing facility had scaled output of its third-generation Figure 03 robot from one unit a day to one unit an hour - a 24-fold increase in throughput in under 120 days, with more than 350 units produced and over 9,000 actuators manufactured across ten-plus SKUs.^[1] That is not a capability claim. It is a manufacturing metric, audited the same way any other production line is audited.

The same shift shows up in how these companies are now being asked to prove themselves publicly rather than on their own terms. Figure livestreamed its Figure 03 humanoid sorting packages continuously for over 163 hours, unedited, reportedly completing more than 204,000 sorts in that window.^[2] Waymo's safety claims are not company marketing - they are drawn from data submitted to the California Public Utilities Commission and the NHTSA's Standing General Order reporting programme, the same regulatory channel every other AV operator has to file into.^[3] Unitree's profitability figures are sitting in a prospectus that China's securities regulator has now approved for public listing, subject to the disclosure standards that come with that.^[4] None of this is a video anymore. It is the kind of evidence that gets audited, regulated, and priced.

2. Four Different Bets on What "Acting" Means

What is striking about 2026 is not that one company has cracked physical AI - it is how differently the leading players are defining what "acting" should even mean. At least four distinct bets are running in parallel, and they are not converging on a single approach.

Bodies that do many things, in one place. Figure AI's bet is the humanoid generalist: one robot, trained on its own Helix vision-language-action model rather than a licensed one, learning tasks from demonstration video and applying them across a live production environment.^[1] Boston Dynamics is running a version of the same bet from a very different ownership position - its all-electric Atlas humanoid had its entire 2026 production run committed before the year began, split between parent company Hyundai's own Robotics Metaplant Application Center and a research partnership with Google DeepMind, which supplies Atlas's Gemini Robotics foundation models as its reasoning layer.^[5] Hyundai does not just buy Atlas - it owns Boston Dynamics outright, which means the roadmap for the robot's deployment is set by manufacturing strategy rather than negotiated through a sales contract.^[5]

A brain for many different bodies. Physical Intelligence is making close to the opposite bet: rather than one company controlling both the robot and the model, its π-series models are designed as a control layer that can, in principle, sit on top of anyone's hardware. Its April 2026 model, π0.7, was the first to show what the company calls compositional generalization - combining skills learned in separate contexts to attempt tasks it had never been trained on directly.^[6] Co-founder Sergey Levine was careful to frame this as an early research result rather than a deployed capability, and the company's own researchers have been candid that a chunk of their apparent failures came down to poor task instructions rather than model limitations.^[6] That candour is itself a useful signal - this is a field where the honest players are the ones worth watching most closely.

Acting at civilisation scale, one ride at a time. Waymo's version of Act has nothing to do with humanoids. It is a fleet of roughly 3,700 to 4,000 vehicles delivering close to half a million fully driverless rides a week across more than ten US cities and over 1,400 square miles of service area - larger than the state of Rhode Island - with four more cities added in a single announcement in July 2026.^[3] Waymo's own safety data reports meaningfully fewer serious-injury crashes than human drivers over the same distance, though the company's own executives have been explicit that its stated target of a million weekly rides by year-end is a stretch goal it may not fully reach.^[3] This is Act as a transport network, not a robot.

Capital markets pricing Act directly. Unitree's answer to "what does acting look like" is the least glamorous and arguably the most consequential: ship more units than anyone else, at a lower cost, and be profitable doing it. The company shipped over 5,500 humanoid and quadruped robots in 2025 alone - more than the combined output of Tesla, Figure, and Agility Robotics - posted a 674% year-on-year jump in net profit, and had its Shanghai IPO application approved by China's securities regulator in July 2026 in the fastest review cycle the exchange's pre-review process has recorded.^[4] An industrial-grade embodied model is already running pilot deployments inside Unitree's own factories, autonomously completing joint motor assembly, with clients including State Grid, PetroChina, and Amazon.^[7] The honest caveat, buried in its own prospectus: 73.6% of humanoid revenue still comes from research and education customers rather than industrial deployment, so the shipment numbers are running well ahead of the "acting in a real workplace" numbers.^[7]

Tesla belongs in this list too, if only because it is impossible to write about Act in 2026 without addressing it - but it deserves a more cautious sentence than the others. Elon Musk has stated that Optimus Gen 3 production begins in summer 2026 and that over a thousand earlier-generation units are already working Tesla's own factory floor, positioning it as the highest-volume humanoid deployment anywhere.^[8] Independent reporting is considerably less certain: several outlets note that as of mid-2026 Optimus has not yet been confirmed doing sustained productive work outside Tesla's own facilities, that production timelines have slipped before, and that credible robotics researchers have publicly and sharply disputed how close the platform is to general-purpose usefulness.^[8] Of everything in this article, Optimus is the entry where the gap between the claim and the independently verifiable evidence is widest - worth watching, not yet worth taking at face value.

3. The Bridge Companies

Two companies that get grouped into this conversation - Prometheus and Odyssey - are worth naming precisely because they are not, strictly, Act companies at all, and the distinction matters. Jeff Bezos has been explicit that Prometheus, the AI startup he co-leads which raised $12 billion in mid-2026 at a $41 billion valuation, has "nothing to do with robotics."^[9] What it is building is an "artificial general engineer" intended to compress the design-to-manufacture cycle for physical products - Bezos's own example is a jet engine redesign that currently takes a decade, which Prometheus wants to turn into an end-to-end AI problem solvable many times faster.^[9] That is AI accelerating how physical things get built, not AI executing physical action - closer to a hybrid of Sense and Reason applied to engineering itself.

Odyssey sits even more squarely in Reason. Its own positioning is unambiguous: it is "pioneering general world models: causal, multimodal systems that learn to understand and simulate the world," with models like Odyssey-2 generating minutes-long interactive video simulations frame by frame.^[10] Odyssey itself does not actuate anything - but its stated purpose is explicitly to accelerate robotics downstream, on the argument that a world model which understands cause and effect, physics, and human interaction gives a robot something a purely visual model cannot: the ability to adapt to a situation it has never encountered rather than merely pattern-match to one it has.^[10] That is precisely the argument I made in the world models article in this series, and it is not a coincidence that Odyssey is now the company most often cited as the model for what that article described in the abstract.

There is a genuinely enjoyable New Zealand thread running through this part of the story, worth a brief aside. Odyssey's CTO, Jeff Hawke, is an Auckland-raised engineer whose first job out of university was writing algorithms for an autonomous forklift - a fitting origin story for someone now building world models that other Physical AI companies train against.^[12] A second New Zealand-founded company sits right next to Odyssey in this bridge layer: Antioch, a simulation platform co-founded by Harry Mellsop, which lets robotics teams build, test, and validate autonomous systems entirely in software before any hardware leaves the lab. Antioch's pitch is that most robotics teams are still testing the expensive way - renting warehouses, resetting hardware between runs, in some cases literally renting Airbnbs to trial household robots overnight - and that the same simulation discipline Tesla and Waymo built in-house at enormous cost should be available to any autonomy team as a platform.^[13] I have connected with both founders recently, and it is a reminder that the infrastructure layer underneath this entire Act pillar - not just the headline robots - is where some of the most interesting New Zealand-connected activity in Physical AI is actually happening.

NVIDIA occupies a similar bridging role from the infrastructure side rather than the model side. Its own framing of physical AI rests on three pillars - simulation environments like Omniverse and Isaac Sim for generating synthetic training data, foundation models like Project GR00T for humanoid control, and onboard compute like Jetson Thor for real-time inference - and in mid-2026 it released a wave of open-source tooling aimed specifically at reducing the cost and complexity of moving a model from simulation into a physical robot.^[11] None of these four companies act. All four exist because the gap between a model that reasons well in simulation and a robot that acts reliably in the physical world is still the hardest part of this entire field to close - which is exactly why they are worth watching as closely as the robots themselves.

4. What "Acting" Still Doesn't Mean

It would be easy to read the numbers above and conclude the problem is solved. It is not, and the companies with the most credibility in this space are the ones saying so themselves. Physical Intelligence's own paper on π0.7 uses careful, hedged language throughout - "early signs" of generalization, "initial demonstrations" - and its co-founder declined to speculate on when the research would become a deployed product.^[6] Unitree's prospectus discloses, in its own words, that its general embodied models "have not yet been applied at scale to robot products," a striking admission from the sector's highest-volume shipper to make in the document underpinning its own IPO.^[7] MIT roboticist and iRobot co-founder Rodney Brooks has publicly called the vision of humanoid robots as general-purpose catchall assistants "pure fantasy thinking," citing coordination challenges he does not believe current technology has actually solved.^[8]

There is also a distinction worth holding onto between teleoperation, supervised autonomy, and genuine unsupervised autonomy, because press coverage routinely blurs the three together. A robot operating on a fenced, lit, pre-measured production cell - the traditional strength of industrial automation - is doing something categorically easier than a robot operating in an unstructured warehouse, and a robot doing either of those is doing something categorically easier than one operating safely around the public with no fallback human in the loop at all. Waymo has spent over a decade and 220 million driverless miles narrowing that gap for one use case.^[3] Most of the humanoid deployments described above are still, in various ways, closer to the fenced-cell end of that spectrum than the coverage often suggests.

5. The Progress Report, in One Line

The reason this matters for the Synaptec Physical AI and AIoT Framework is not that any of this is finished - it is that Act has, for the first time, become a category generating its own independent evidence rather than borrowing credibility from Sense and Reason. A model that reasons well is not the same claim as a robot that acts reliably, and 2026 is the year enough companies made the second claim, and were tested on it publicly enough, that the distinction stopped being theoretical. That is the honest state of the progress report: real, verifiable, still early, and moving faster than most governance conversations have caught up to.

That evidence cuts both ways. Every one of the metrics in this article - production throughput, ride volume, IPO prospectuses, livestreamed uptime - exists precisely because these systems are now doing something in the physical world with real consequences if they get it wrong. A robot sorting 200,000 packages autonomously is remarkable. It is also 200,000 opportunities for something to go wrong on an active production floor, with no human confirming each individual action. The same acceleration that makes this article exciting to write is what makes the accountability question unavoidable - who is responsible when one of these systems acts and gets it wrong, under what oversight model, and covered by whose insurance. That is the subject of next month's piece in this series.