Generalist AI

Overview

Generalist AI builds foundation models that control robot arms and hands. Its models take in video, sensor readings and language and emit motor commands in real time, and they are meant to run on robots the company does not make. What separates it from other robot model labs is the data. Instead of paying operators to drive robots in a lab, Generalist sends cheap wearable grippers to people around the world, who use them to do ordinary manual tasks, and it trains on the recordings. The company says the resulting dataset passed half a million hours in April 2026, which on the public numbers is the largest real-world manipulation dataset any lab has disclosed. It has used that data to publish the first scaling laws for robot pretraining and three model generations in ten months.

Total funding combines the company’s statement of more than $500 million raised as of June 2026 with an extension of about $200 million reported by Axios and TechCrunch in August 2026. The employee count is from Tracxn as of August 31, 2026, and is not company-confirmed.

Thesis

Language models became general because the internet had already written down trillions of words. Robotics has no equivalent. There is no archive of the forces, contacts, corrections and small recoveries involved in folding a box or threading a cable, and a video of a person doing it records the pixels but not the physical interaction. The Daily Upside put the problem plainly when it wrote that the dataset for physical interaction does not really exist. Every robot model company is therefore, first, a data company, and what separates them is which collection method produces useful signal fastest and cheapest.

The dominant method until recently was teleoperation, in which a person drives a real robot through a task while the system records what happens. It produces clean data for that robot, but it is slow and expensive, since each operator needs a robot, and the robot has to be maintained. Physical Intelligence’s first model, π0, was trained on about 10,000 hours of robot experience. Nvidia’s GR00T N1 paper describes expanding 88 hours of in-house teleoperation to 827 hours with generated variations and filling the rest of its mixture with simulation and human video. At those volumes, robot learning looked like a field that would advance through clever algorithms rather than through scale.

Generalist’s bet is that the language model recipe does carry over once the data problem is solved, and that the fastest way to solve it is to separate data collection from robots entirely. Its GEN-0 post in November 2025 reported 270,000 hours of manipulation data growing at 10,000 hours a week, gathered across thousands of homes, warehouses and workplaces, and showed validation error falling as a power law in dataset size. It also reported something that looks like a threshold. Models of 1 billion parameters stopped absorbing new data, a failure the company calls ossification, while models of 7 billion parameters and above kept improving. If that result holds, robot learning has the same property that made language models predictable, which is that more data and more compute buy capability in a way a lab can plan around.

The commercial argument follows from that. If one model can drive many robot bodies, and if new tasks can be taught in minutes rather than engineered in weeks, then the value in robotics moves from hardware and custom integration toward whoever owns the model. The installed base is large. The International Federation of Robotics counted 4.66 million industrial robots in operation worldwide in 2024, with 542,000 installed that year, most of them programmed for a single fixed job. A model that let those arms handle variable, high-mix work would open a market that conventional automation has never served economically.

Founding Story

Generalist was founded by Pete Florence, Andy Zeng and Andrew Barry. TechCrunch and Tracxn give the founding year as 2024, while the notes to Florence’s August 2026 podcast appearance with Joe Lonsdale say he left Google in 2023 to co-found the company, so work likely began before incorporation.

Florence, the chief executive, was a senior research scientist at Google DeepMind, where he helped lead PaLM-E and RT-2. PaLM-E was an early demonstration that a large language model could be extended to take in camera images and robot state, and RT-2 showed that a vision-language model could output robot actions directly, the architecture now called a vision-language-action model. Zeng, the chief scientist, earned his PhD at Princeton and was a staff research scientist and tech lead at Google DeepMind. His prior work includes TossingBot, a system that learned to throw arbitrary objects, Code as Policies, which used language models to write robot control code, and PaLM-E. Barry, the chief technology officer, earned his PhD in MIT’s Robot Locomotion Group working on high-speed autonomous drone flight, where Florence was a classmate, and then worked at Boston Dynamics, including on the Spot robot. A post on the More Entropy newsletter, written by an investor who backed the company early, says the founding team also included Evan Morikawa, who previously worked on product development at OpenAI.

The combination matters for the thesis. Florence and Zeng were among the researchers who showed that internet-scale models could be connected to robots at all, and they left Google with a clear view of what those models lacked, which was physical data. Barry brings the hardware side needed to build the collection devices and the robot stack. Florence has described the founding insight as the recognition that intelligence, more than hardware, is robotics’ next frontier, an idea he says came from watching lab robots that sat idle most of the day for lack of anything useful to do.

Early backers included Spark Capital, Boldstart Ventures, Nvidia’s NVentures, Bezos Expeditions and NFDG, the fund of Nat Friedman and Daniel Gross. Tracxn records a seed round in March 2025 led by NVentures, with the amount undisclosed. The company stayed quiet until the GEN-0 release in November 2025.

Product

Generalist sells, or intends to sell, a model rather than a robot. The company buys commercially available arms and hands for its own work and keeps its engineering focused on the model and the data. Its product line to date is three model generations, the data engine that feeds them, and the collection hardware that feeds the data engine.

Data hands

The collection device is a strap-on pincer that a person wears on the hand and uses to do real tasks, closing and opening it to grasp objects. Humanoids Daily describes them as low-cost devices in the family of universal manipulation interfaces, a design popularized in academic research, and reports that they capture reflexes and small corrections more efficiently than latency-prone teleoperation rigs. The Daily Upside reports that the company has sent thousands of them to people around the world, who use them for tasks such as tying zip ties and opening bags of chips. Because the grippers approximate the robot’s own end effector, the recorded motions can be mapped onto a robot more directly than footage of bare human hands.

A July 2026 post, Towards Machines with a Thousand Hands, extends the idea beyond two-finger grippers. It describes a pretraining dataset spanning roughly 9,000 variations of end effector, from five-finger hands to power screwdrivers, tongs, whisks, peelers and box cutters. The company has not disclosed what a device costs, how many are in circulation, or how collectors are paid. The GEN-0 post refers to several data foundry partners with different operating approaches, which suggests that at least part of the collection is contracted out.

Data engine

The GEN-0 post describes infrastructure that processes dozens of petabytes of compressed data on about ten thousand processor cores and absorbs 6.85 years of manipulation experience per day of training, figures that are company-reported. At this scale the bottleneck moves from collecting data to cleaning and loading it, the same shift language model labs went through.

GEN-0

GEN-0, released in November 2025, is the research foundation. It was trained on 270,000 hours of data, scaled past 10 billion parameters, and tested on robots with 6, 7 and more than 16 degrees of freedom, the last a semi-humanoid. Its main contributions were the scaling results already described, the 7 billion parameter threshold, and an architecture feature the company calls Harmonic Reasoning, which lets the model continue to reason while it is acting rather than alternating between thinking and moving. Post-training on as little as 5.6 hours of task-specific data produced large gains, and the company reported real-robot success of up to 99% in some cases.

GEN-1

GEN-1, released on April 2, 2026, is the model Generalist presents as crossing into commercial viability. It was pretrained on more than 500,000 hours of human wearable data with no robot data at all, and then adapted to each task with about one hour of robot-specific data. On the tasks the company chose to report, including folding boxes, packing phones and servicing robot vacuums, average success rose to 99% from 64% for GEN-0 and 19% for a model trained from scratch. Box folding took 12.1 seconds against about 34 seconds for GEN-0 and for Physical Intelligence’s π0 and π*0.6 models, which is the basis of the claim of up to three times prior state-of-the-art speed. The company also reported long unbroken runs, including 1,800 consecutive block packing cycles, more than 200 vacuum services, 86 T-shirt folds in a row, and more than an hour of kitting auto parts without intervention.

The post also highlighted emergent behavior, such as a robot using its second arm to shake a bag when a toy snagged inside it, and was candid that GEN-1 is far from perfect and that improvisation can produce unintended physical consequences. The company launched an initiative it calls alignment for embodied intelligence alongside the model. GEN-1 was made available only to early access partners.

GEN-1.5

GEN-1.5, released on August 19, 2026, changes the interface. Instead of post-training on an hour of data, a user shows the robot a single demonstration of 3 to 12 seconds, and the model performs the task with no change to its weights, which the company calls physical prompting. The model processes video with sensor, language and joint data, holds 30 seconds of memory in context, and outputs actions at 100 Hz. It was trained continuously for more than eight months on over 500,000 hours of data and used no simulation data.

The reported numbers are modest and the company says so. Across 10 short manipulation tasks such as zippers, jars and wallets, one-shot prompting succeeded 59% of the time on average, with a standard deviation of 10 points. Ten gradient steps on five minutes of data per task, about 50 demonstrations, raised that to 83%. The post describes the tasks as simple and short-horizon and says in-context skills are more brittle than fine-tuned ones. The significance is less the success rate than the cost of teaching. If a task that used to require days of programming can be set up by a person demonstrating it once, the economics of automating low-volume work change even at modest reliability, provided a fine-tuning step closes the gap.

Customer

Generalist has not named a customer. TechCrunch reported in August 2026, citing sources, that the company is working with a handful of customers and uses their feedback to tailor models to specific use cases. GEN-1 was released to early access partners and GEN-1.5 to design partners, and the company’s own pages direct interested parties to a partnerships address rather than to a product page or price list. The careers page lists 23 open roles across San Francisco and Boston, of which three, a partnerships associate, a partnerships lead and a strategic deals lead, are commercial.

The tasks the company chose to demonstrate indicate the target buyer. Kitting auto parts, folding boxes, packing phones and servicing appliances are the repetitive, variable, hand-dependent work found in light manufacturing, electronics, logistics and returns processing. That work is still largely done by people because each variation would need its own conventional robot program, and the volume per variation is too low to justify one. Sacra, a third-party research firm, describes the target as industrial operators, manufacturers and logistics companies.

One hardware relationship has surfaced from the other side. In July 2026, the Chinese cobot maker Elite Robots said its arms were the hardware underneath some of GEN-1’s demonstrations and had been used for data collection and validation. Unite.AI noted that Generalist’s own posts never mention Elite, so the relationship is confirmed only by the supplier.

Market Size

Nobody measures the market for robot foundation models directly, because it barely exists as a revenue category. The available figures describe adjacent or much broader markets, and they differ by orders of magnitude depending on the definition.

All three dollar figures are third-party forecasts. The Goldman number is the most revealing about how fast expectations are moving. In February 2024 the bank projected a $38 billion humanoid market by 2035 on 1.4 million units, and in September 2026 it raised that to about 6.5 million units and roughly $138 billion, almost a fourfold revision in two and a half years. The Research and Markets figure counts autonomous vehicles, drones, smart infrastructure and other systems, so it describes the broad theme rather than anything Generalist sells into.

A bottom-up frame is more useful. A model company’s revenue scales with the number of robots running its model and the fee per robot. The 4.66 million industrial robots in operation, plus 542,000 added each year, are the hardware a model could be retrofitted to, and forecast humanoid fleets would add to it. No one has published either the fee per robot or the share of that base doing variable work.

The capital flowing in is measured more reliably than the revenue. Crunchbase reported that robotics startups raised about $15 billion in 2025, passing the 2021 peak, and $18.8 billion in 2026 by June 22, more than any prior full year. An earlier Crunchbase article gave the 2025 figure as nearly $14 billion, a 70% rise over 2024, so the database’s totals shift as rounds are added.

Competition

Generalist competes with three kinds of company. Independent robot model labs sell the same layer. Humanoid makers build their own models for their own hardware. Large platforms publish models that may set the price of the layer near zero.

Independent robot model labs

Physical Intelligence is the closest rival in thesis, team and funding. It was founded by academics and former Google DeepMind researchers, including Sergey Levine and Chelsea Finn. According to its Wikipedia profile, it raised $400 million at $2.4 billion in 2024 and $600 million at $5.6 billion in 2025, and Bloomberg reported in March 2026 that it was in talks to raise about $1 billion at more than $11 billion. Its π0 model was open-sourced, and its π*0.6 model, released in November 2025, uses reinforcement learning from the robot’s own experience and corrections, and ran for 13 hours making espresso drinks in the company’s office. The difference with Generalist is mainly in data. Physical Intelligence has relied on robot teleoperation and on learning from deployment, while Generalist has relied on human wearable data at much larger volume. Public profiles describe Physical Intelligence’s deployments as tests, such as folding laundry in short-term rentals, and say it does not yet sell its technology commercially.

Skild AI is the rival with a commercial lead. It raised a $1.4 billion Series C led by SoftBank in January 2026 at more than $14 billion, has raised more than $2 billion in total, and told Bloomberg in September 2026 that it had reached a $100 million annual revenue run rate ten months after beginning commercial use, with its software on hundreds of robots at more than 60 companies, up from eight earlier in the year. Those figures are company-reported. Its S1 model, announced in August 2026, learns long-horizon tasks from one human video and reports 66% success on unseen tasks against 9% for a language-prompted baseline, with humans recovering failed steps during grading. GEN-1.5 and S1 appeared within days of each other and make nearly the same claim, which suggests that one-shot demonstration is where the frontier currently sits and that neither lab holds it alone.

Dyna Robotics is smaller but further along in production. It raised a $120 million Series A in September 2025, about $143.5 million in total, at a valuation above $600 million. In August 2026 it described taking a commercial pilot with the restaurant chain Din Tai Fung through to scaled deployment, with its Dyna-2 model folding 95 napkins an hour at 93% quality and hundreds of robots expected by the first half of 2027. It narrows to one task done profitably, while Generalist widens toward general capability first.

Genesis AI builds both a model and its own dexterous hands. It raised a $105 million seed in 2025 and was reported by Bloomberg in July 2026 to be in talks to raise about $500 million at about $3 billion. Field AI, founded by NASA JPL veterans, raised about $405 million across 2025 at a $2 billion valuation and focuses on autonomy for field robots in construction, energy and defense rather than dexterous manipulation. Rhoda AI raised $450 million at $1.7 billion, per The Daily Upside, on a strategy of training on public online video, a cheaper data source than Generalist’s but one that captures no physical interaction.

Humanoid makers with in-house models

Figure AI raised more than $1 billion at a $39 billion post-money valuation in September 2025 and develops its own vision-language-action model, Helix, having ended its model collaboration with OpenAI in February 2025. The company says its Figure 02 robot supported production of more than 30,000 BMW X3 vehicles over ten months. Tesla is building Optimus with its own AI stack derived from its driving software. Neither is likely to buy a model, because each treats the model as the core of its product. If humanoids dominate and their makers keep models in house, an independent lab’s market narrows to arms, mobile manipulators and humanoid builders without model teams.

Platforms and open models

Google DeepMind, where Generalist’s founders did their earlier work, publishes the Gemini Robotics family, including an on-device version that runs locally on the robot. Nvidia publishes the open GR00T models for humanoids alongside its simulation tools. Nvidia is also an investor in Generalist, Skild, Figure, Dyna and Field AI, so its interest is in the category growing on its hardware rather than in any one lab winning. Open models from either company set a free baseline that paid models must clearly beat, much as open-weight language models did for text.

Business Model

Generalist has published no pricing and no revenue. Sacra describes its current commercial approach as direct engagement through pilot fees and deployment contracts rather than self-serve subscriptions, which is typical for a company whose product has to be validated on each customer’s floor before anyone pays for it at scale. That description is third-party and not confirmed by the company.

Over time, a robot model company can charge per robot running the model, charge for usage such as cloud inference or teaching new tasks, or take a share of the labor value through robots as a service with a hardware partner. Skild’s disclosures suggest per-customer software revenue is achievable today. Its $100 million run rate across more than 60 companies implies an average above $1.5 million a year per customer, though concentration and pilot structure are unknown.

The table below shows the arithmetic of the first model with an illustrative fee of $10,000 per robot per year. The fee is an assumption for scale, not a figure Generalist or any rival has published.

A hundred thousand robots would be a little over 2% of today’s industrial installed base. The table is a reminder that a multibillion-dollar valuation for a pure model company implies either very broad adoption or a fee far above the cost of a conventional robot program.

The cost side is heavy. The company pays for data collection hardware and collectors, for data foundry partners, for a data engine that processes petabytes, and for the compute to train models of more than 10 billion parameters continuously for eight months at a time. It also has to support deployments until they run unattended. The advantage of the wearable approach is that the marginal hour of data is far cheaper than an hour of teleoperation, and the data is not tied to any one robot. The risk is that the largest cost, training compute, scales with ambition rather than with customers.

Traction

Technical figures are company-reported. The customer description is from TechCrunch, citing sources.

The traction that can be checked is research output and capital. Generalist has shipped a major model roughly every four months, each with a detailed technical post, and each one extended the previous result rather than repeating it. GEN-0 showed scaling, GEN-1 showed reliability after a small amount of robot data, and GEN-1.5 showed learning from a single demonstration. The posts are unusually forthcoming about limitations, including GEN-1.5’s own description of its success rates as modest.

The capital came quickly. The company announced a $400 million round on June 4, 2026, led by Radical Ventures with 8VC, Union Square Ventures, Hanabi Capital and Norwest as new investors, existing investors NVentures, Boldstart, Spark Capital, Bezos Expeditions and NFDG returning, and angels including Fei-Fei Li, Xiaomi co-founder Lin Bin, Naval Ravikant and Zoom’s Eric Yuan. Bloomberg reported the valuation at $2 billion. Less than three months later, Axios reported an additional $200 million, and TechCrunch reported that the extension, led by 8VC, valued the company at $3 billion. The company has not announced the extension itself. Florence has said, per the Lonsdale podcast notes, that no technical employee has left the company in two and a half years, which if accurate is notable in a field where rival labs recruit aggressively.

What does not yet exist in public is commercial traction. There is no revenue figure, no named customer, no deployment count and no case study. Measured against Skild’s revenue and customer disclosures, or against Dyna’s documented restaurant rollout, Generalist is behind on everything a buyer of robots would check.

Valuation

Generalist’s announced valuation is $2 billion, set by the June 2026 round, and its reported valuation is $3 billion after the August extension. At $3 billion on roughly $700 million raised, investors own a large share of the company for this stage, which reflects how capital-intensive the strategy is.

The comparable marks cluster in two groups. The leaders are Skild at more than $14 billion with a $100 million run rate, which works out to roughly 140 times annualized revenue, and Physical Intelligence at a reported $11 billion with no disclosed revenue. The next tier includes Generalist at $3 billion reported, Genesis in talks at about $3 billion, Field AI at $2 billion and Rhoda at $1.7 billion. Figure, at $39 billion, is priced as a humanoid manufacturer rather than a model company.

The gap between Generalist and Physical Intelligence is the most informative. The two have related founding teams and the same thesis, and Generalist has published comparable technical results and a larger dataset, yet Physical Intelligence is priced at more than three times as much. That gap reflects its earlier start, larger capital base and visibility more than any disclosed result. The Skild multiple sets the other boundary. Investors pay about 140 times revenue for the one company in the category with revenue, so a lab without revenue is priced almost entirely on expected future share.

Key Opportunities

A data flywheel that compounds

Once models are useful, deployed robots generate data of their own. A company that pairs a large human data corpus with a growing fleet of robot data from customer deployments would have two sources compounding at once. Generalist’s scaling results suggest each additional hour is worth something predictable, which turns data collection into an investment decision rather than a research gamble.

Programming by demonstration as a product

GEN-1.5’s physical prompting, if it matures, is a product rather than a research result. Most industrial robots today are programmed by specialists or integrators, a process that can take days per task and is the main reason high-mix, low-volume work stays manual. A model that lets a line worker teach an arm a new task by doing it once would sell to the installed base of arms already in factories, without new hardware.

Becoming the default model for arm makers

Elite Robots publicly positioning its cobots as hardware for GEN-1 shows the shape of this opportunity. If arm and hand manufacturers compete to be compatible with a leading model, the model company gains distribution without building a sales force for each factory, much as Android spread through handset makers.

Key Risks

No commercial evidence yet

This is the strongest bear case. Generalist has no disclosed revenue, no named customer and no deployment count, while Skild reports a $100 million run rate across more than 60 companies and Dyna has a documented restaurant rollout. Robotics has a long history of impressive lab results that did not survive contact with real factories, where lighting, part variation, uptime requirements and safety certification all bite. Until Generalist shows a customer running its model unattended at a profit, its commercial viability rests on its own description.

Self-reported benchmarks on self-chosen tasks

The 99% figure is an average across tasks the company selected, after task-specific post-training, measured on its own setups. There is no shared, independent benchmark for robot manipulation comparable to the benchmarks used for language models, and every lab reports on its own tasks. The comparison to Physical Intelligence’s box folding speed is a comparison to a competitor’s published video, not a head-to-head test. None of the results has been reproduced by an outside party.

Human data may not transfer to every robot

GEN-1 was pretrained entirely on data from people wearing grippers. That works when the robot’s hand resembles the gripper, and the company is expanding to thousands of end effector variations to broaden it. But the robots customers already own have their own kinematics, payloads, speeds and failure modes, and a human arm moves differently from a six-axis industrial arm. If each new robot type needs substantial adaptation, the advantage of robot-independent data shrinks. Critics quoted by Humanoids Daily have questioned the approach from other angles as well. Cobot’s chief executive Brad Porter called data-heavy approaches really expensive without the right architecture, and Yann LeCun has argued that world models must learn from observation rather than from predicting action tokens.

Better-funded rivals with the same thesis

Physical Intelligence has raised more capital, has a similar founding team and a similar goal, and has pursued reinforcement learning from deployment as an alternative route to reliability. Skild has more capital still, a commercial lead and a one-shot model announced within days of GEN-1.5. In a field where training compute is the largest cost, a rival with several times the capital can train larger models and subsidize early deployments longer.

Hardware makers keep the model in house

Figure and Tesla develop their own models, and other humanoid makers may follow if models become the main differentiator. Nvidia and Google DeepMind publish capable models that hardware makers can adopt without paying a license. If the robots that matter most are built by companies that do not buy models, an independent lab is left selling into the rest of the market on price.

Valuation ahead of revenue across the category

Robotics venture funding in the first half of 2026 exceeded every prior full year. The category’s valuations assume a general-purpose robot market that forecasters themselves keep revising, as Goldman’s near fourfold change shows. A slowdown in funding before revenue arrives would hurt capital-intensive labs with no revenue first, and Generalist is one of them.

Summary

Generalist AI trains foundation models for robot manipulation on data collected by people wearing low-cost grippers, rather than by operators driving robots. The approach has produced the largest real-world manipulation dataset any lab has disclosed, more than 500,000 hours, and the first published scaling laws for robot pretraining. The founders include the Google DeepMind researchers behind PaLM-E and RT-2 and a Boston Dynamics engineer, and the company has released three model generations in ten months, each with a detailed technical account.

As of September 2026, Generalist had raised about $700 million, including a $400 million round led by Radical Ventures at $2 billion in June and a reported extension of about $200 million led by 8VC at $3 billion in August. Its GEN-1 model reports 99% average success on selected tasks after one hour of robot data, and GEN-1.5 reports 59% success on simple tasks from a single short demonstration, rising to 83% with a few minutes of fine-tuning. It has about 106 employees, according to Tracxn, in San Francisco and Boston offices.

The central question is whether a data advantage turns into a commercial one. Generalist’s technical record is at least as strong as that of rivals priced several times higher, but its commercial record is thinner than Skild’s, which already reports $100 million in run-rate revenue, and thinner than Dyna’s, which has a documented deployment. Its human data has to transfer to the robots customers actually own, its one-shot results have to reach factory reliability, and it has to find buyers before better-funded rivals and hardware makers with their own models fill the layer it wants to own.

The technical claims above are company-reported and none has been independently reproduced. The company has disclosed no revenue, no pricing and no customer, and the $3 billion valuation and the $200 million extension come from press reports rather than from the company.