Generalist AI
Last reviewed
Jun 4, 2026
Sources
17 citations
Review status
Source-backed
Revision
v1 · 1,685 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 4, 2026
Sources
17 citations
Review status
Source-backed
Revision
v1 · 1,685 words
Add missing citations, update stale details, or suggest a clearer explanation.
Generalist AI (branded Generalist, legally Generalist AI, Inc.) is an American artificial intelligence company building general-purpose foundation models for robots and other physical systems. Founded in 2024 and based in San Mateo, California, the company was started by former Google DeepMind robotics scientists Pete Florence and Andy Zeng together with former Boston Dynamics engineer Andrew Barry. Generalist trains a single family of embodied AI models intended to run across many different robots, from humanoids and warehouse machines to industrial arms, rather than building any one robot of its own. In June 2026 the company raised $400 million at a reported $2 billion valuation in a round led by Radical Ventures, with Nvidia and Bezos Expeditions among its backers, bringing total funding to more than half a billion dollars.
The company describes itself as "a frontier AI research and product driven company building general intelligence for the physical world," with the tagline "We train robots that work." Its public models are named in the GEN series: GEN-0, released in November 2025, and GEN-1, released in April 2026.
Generalist was founded in 2024 by three researchers with deep backgrounds in robot learning. Pete Florence, the company's chief executive, earned a PhD in computer science from MIT under Russ Tedrake and then worked as a senior research scientist at Google DeepMind, where he helped create RT-2 (a vision-language-action model that transfers web knowledge to robotic control) and PaLM-E (an embodied multimodal language model). Andy Zeng, the chief scientist, studied at UC Berkeley and earned a PhD at Princeton, then worked as a research scientist and tech lead at Google DeepMind, where he contributed to PaLM-E, to methods for robots that "write their own code," and to low-cost handheld data collection for robots. Andrew Barry, the chief technology officer, was a roboticist at Boston Dynamics and was Florence's PhD classmate at MIT. The broader team includes engineers from OpenAI, Google DeepMind, and Boston Dynamics.
The startup spent its first year largely in stealth. It first drew public attention in March 2025, when TechCrunch reported that Florence had left DeepMind and that Nvidia's venture arm had already backed his new company. At that stage Florence framed the mission as making general-purpose robots a reality and ultimately driving "the marginal cost of physical labor to zero." Florence's DeepMind research was cited several times in DeepMind's own March 2025 robotics paper, underscoring how recently he had left frontier robotics research to start the company.
Generalist's seed funding came together in 2024. Boldstart Ventures records its first investment in the company on March 24, 2024, alongside Nvidia as a co-investor, and Boldstart, Spark Capital, and NFDG were among the early backers. Public databases also list a seed round dated to March 2025, reflecting the period when the previously stealth company became more visible.
On June 4, 2026, Generalist announced a $400 million funding round, reported by Bloomberg at a $2 billion valuation. Radical Ventures led the round, with new investors 8VC, Union Square Ventures, Hanabi Capital, and Norwest, and returning investors including Nvidia's NVentures, Boldstart Ventures, Spark Capital, Bezos Expeditions, and NFDG. Angel investors in the round included Zoom founder Eric Yuan, Xiaomi co-founder Bin Lin, AI researcher Fei-Fei Li, and entrepreneur Naval Ravikant. The round brought Generalist's total funding to more than $500 million. Some outlets characterized the financing as a Series B, while others described it only as a new round; the company said it would use the capital to build new models, scale its physical-data engine, expand compute and training infrastructure, and pursue commercialization partnerships.
| Date | Event | Amount | Lead / notable investors |
|---|---|---|---|
| March 2024 | Seed (first check) | Undisclosed | Boldstart Ventures, Nvidia, Spark Capital, NFDG |
| June 4, 2026 | New round (reported Series B) | $400 million | Radical Ventures; Nvidia NVentures, Bezos Expeditions, 8VC, Union Square Ventures, Norwest, Hanabi Capital |
Reported valuation at the June 2026 round was about $2 billion, and cumulative funding exceeded $500 million.
Generalist pursues a strategy that differs from most well-funded robotics companies, which tend to design a specific machine such as a humanoid or a warehouse robot. Instead, Generalist concentrates on the software, training one cross-embodiment foundation model meant to control many robot types. The company frames this around the idea that "the future of robotics is bigger than any single robot," and its public materials stress that "goals are more powerful than methods." It has positioned its models as distinct from both conventional vision-language-action models (VLAs) and from world-model approaches, saying GEN-1 is trained from scratch rather than fine-tuned from an existing language or vision backbone.
Central to the approach is a large, proprietary dataset of real-world physical interaction. The company collects manipulation data across thousands of homes, warehouses, and workplaces worldwide, in part using wearable "data hands" (UMI-style handheld grippers) that capture human reflexes and micro-corrections. By the time of GEN-0 the dataset stood at roughly 270,000 hours of real-world manipulation data, which the company said was growing by about 10,000 hours per week. This emphasis on scaling real physical-interaction data places Generalist alongside other robot foundation model efforts such as Physical Intelligence and Skild AI, though Generalist's particular claim is that robot performance follows predictable scaling laws as data grows.
GEN-0, announced on November 4, 2025, was presented as a new class of embodied foundation models that scale with physical-interaction data. Generalist reported a power-law relationship between pretraining data and downstream task performance across more than 16 tasks, analogous to the scaling laws seen in large language models. The company also reported what it called the first observation of model "ossification" in robotics: models around 1 billion parameters showed early ossification (an inability to keep absorbing new information), while models of about 7 billion parameters and larger continued to improve as pretraining data grew. Generalist tested GEN-0 across robots with 6, 7, and 16 or more degrees of freedom and discussed model sizes of 1B, 6B, 7B, and 10B-plus parameters.
A defining feature of GEN-0 is what Generalist calls "Harmonic Reasoning," in which the model is trained to think and act at the same time by processing asynchronous, continuous-time streams of sensing and action tokens. Coverage in Andrew Ng's The Batch and elsewhere described the work as showing that training power laws translate from language to robotics. Several outlets framed GEN-0 as a potential "ChatGPT moment" for embodied AI, though that comparison came from commentators rather than from independent benchmarking.
GEN-1, released on April 2, 2026, was described by the company as its first general-purpose model to reach "mastery of simple physical tasks." Generalist reported that GEN-1 raised average success rates to about 99% on tasks where its previous model had averaged roughly 64%, and that it completed dexterous tasks roughly three times faster than prior state-of-the-art approaches. The company said a new robot or task could be adapted with as little as one hour of robot-specific data, because the base model is pretrained primarily on human-demonstration data and only the final stage of training uses the target robot hardware. Training data was reported to have expanded from 270,000 hours for GEN-0 to over 500,000 hours for GEN-1.
Generalist demonstrated sustained reliability on repetitive real-world tasks, citing runs such as 200 consecutive box-folding repetitions, 200-plus robot-vacuum servicing repetitions, and more than 1,800 block-packing repetitions. Florence described GEN-1's behavior in terms of reliability, speed, and "intelligent improvisation," pointing to unrehearsed corrections such as a robot using a second arm to shake a bag when an object snagged. The company said GEN-1 marked a transition from research prototype toward commercial viability and was being made available to early-access partners.
Generalist's emergence has been covered by Bloomberg, The Robot Report, SiliconANGLE, Silicon Republic, and robotics-focused outlets such as Humanoids Daily, and its research has been discussed in DeepLearning.ai's The Batch. Reporting has emphasized the unusually strong investor lineup (Nvidia, Bezos Expeditions, Fei-Fei Li, Naval Ravikant) and the founders' direct lineage to influential DeepMind robotics work like RT-2 and PaLM-E. Observers have placed Generalist within a wave of well-capitalized robot-foundation-model startups racing toward what some describe as "physical AGI." Independent, third-party benchmarking of GEN-0 and GEN-1 remains limited, and most performance figures originate from the company's own reports, a caveat that applies broadly across the robot-foundation-model field.