Genesis AI (company)
Last reviewed
Sources
16 citations
Review status
Source-backed
Revision
v1 · 2,245 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
16 citations
Review status
Source-backed
Revision
v1 · 2,245 words
Add missing citations, update stale details, or suggest a clearer explanation.
Genesis AI is a robotics startup building a universal robot foundation model and a high-fidelity physics-simulation engine to train it. The company describes itself as a "physical AI" research lab and full-stack robotics firm, and it pursues a data-centric strategy: rather than relying only on scarce real-world robot data, it generates large volumes of synthetic training data inside its own simulator. Genesis AI was founded in December 2024 by Zhou Xian, who holds a robotics PhD from Carnegie Mellon University, and Théophile Gervet, a machine-learning PhD from the same school and a former research scientist at the French AI lab Mistral. It emerged from stealth on July 1, 2025 with a $105 million seed round co-led by Eclipse Ventures and Khosla Ventures, and it has offices in both Palo Alto, California and Paris. In May 2026 it unveiled its first model, GENE-26.5, alongside a proprietary dexterous robotic hand, positioning itself among the most heavily funded entrants in the race to build general-purpose embodied AI.
The company shares a name and a founder with the open-source Genesis physics engine, an academic project that Zhou Xian helped lead. The two are related but distinct, and the relationship is explained in the Relationship to the open-source Genesis engine section below.
Genesis AI's central thesis is that progress in robotics is bottlenecked by data. Large language models could be trained on trillions of tokens scraped from the internet, but there is no comparable corpus of physical-interaction data for robots, and collecting it on real hardware is slow and expensive. Genesis AI's answer is a "data-centric, full-stack" pipeline that combines high-fidelity physics simulation, multimodal generative modeling, and large-scale real-robot data collection into a single data engine. Synthetic data produced in simulation is meant to be cheap and effectively unlimited, while real-world data anchors the model to physical reality and helps close the sim-to-real gap.
The company frames the opportunity in macroeconomic terms. In its launch materials it cited an estimate that more than 95 percent of physical labor, representing tens of trillions of dollars of global economic activity, remains unautomated, and CEO Zhou Xian stated that "general-purpose robots powered by physical AI will define the next major chapter of human history." Genesis AI targets domains such as logistics, manufacturing, electronics, pharmaceuticals, and laboratory work, where dexterous, repetitive manual tasks dominate.
Genesis AI was founded in December 2024 by Zhou Xian and Théophile Gervet, who met as PhD students at Carnegie Mellon University. Zhou Xian, the chief executive, completed a PhD at CMU's Robotics Institute focused on robot learning, simulation, and generative models for the physical world. Gervet, who serves as president, completed a machine-learning PhD at CMU under Katerina Fragkiadaki, working on computer vision and 3D robot manipulation, with papers such as Act3D and contributions to embodied-navigation work. Before co-founding Genesis AI, Gervet was a research scientist at Mistral AI, where he worked on open-source model releases; according to reporting by Sifted, he left Mistral in December 2024 to start the company, which was initially associated with the name Genesis Robotics.
The founding team drew researchers from the academic Genesis collaboration as well as alumni of Mistral AI, Nvidia, Google, MIT, Stanford, Columbia, and the University of Maryland. The company says its early staff specialized in robotics, machine learning, and computer graphics.
Genesis AI operated in stealth for roughly six months before its public launch. An earlier report by Sifted in mid-2025 described an $85 million close; the company's formal stealth exit on July 1, 2025 announced a $105 million seed round, one of the largest seed rounds in robotics to date.
| Round | Date | Amount | Lead investors | Other participants |
|---|---|---|---|---|
| Seed | July 1, 2025 | $105 million | Eclipse Ventures, Khosla Ventures | Bpifrance, HSG (formerly Sequoia China), HongShan, Eric Schmidt, Xavier Niel, Daniela Rus, Vladlen Koltun |
The round was notable for combining U.S. and Chinese venture backers at a time when the industry was largely split along geopolitical lines. HSG is the fund formerly known as Sequoia Capital China, and HongShan is its rebranded Chinese arm. Other backers included the French public investment bank Bpifrance and high-profile individuals: former Google chief executive Eric Schmidt, French telecom billionaire Xavier Niel, MIT roboticist Daniela Rus, and computer-vision researcher Vladlen Koltun. As of mid-2026, Genesis AI had not publicly disclosed a post-money valuation or a subsequent priced round.
By May 2026 the company employed roughly 60 people, split about 40 to 45 percent in Europe and 50 to 55 percent in the United States, with offices in Paris, California, and a newly added presence in London. On May 6, 2026 Genesis AI revealed that it had gone "full stack," developing not just software but also its own dexterous robot hand and data-collection hardware, and it launched its first foundation model, GENE-26.5. The launch was accompanied by demonstration videos of a pair of human-scale robotic hands performing manipulation tasks, and the company said it was in late-stage discussions with prospective customers in France, Germany, and Italy.
GENE is Genesis AI's family of robot foundation models. The first release, GENE-26.5, was announced on May 6, 2026 and described by the company as the first "AI brain" to give robots human-level physical manipulation. The naming convention encodes the release timing: "26" refers to the year 2026, and the ".5" denotes the first half of the year (the model shipped in May), with the company signaling that future GENE releases will follow the same year-and-half-year pattern.
The model is designed to be general purpose, controlling different robot hardware across varied environments rather than being tuned to a single platform. In launch demonstrations, robots driven by GENE-26.5 were shown:
These are the company's own demonstrations, and as of mid-2026 the claims of "human-level" dexterity had not been independently benchmarked.
A core part of Genesis AI's full-stack approach is hardware built specifically to gather manipulation data. The company developed a proprietary end effector, a robotic hand that mirrors the human hand in form and function, replacing the simple two-finger grippers common in industrial robot manipulation. It also built a sensor-equipped data-collection glove fitted with tactile-sensing electronic skin. The glove is designed to create a "1:1:1 mapping" between the human hand, the glove, and the robotic hand, so that movements demonstrated by a person transfer directly to the robot without the embodiment gap that complicates other teleoperation setups.
The company says the glove is lightweight, easy to wear, and costs about 100 times less than typical alternatives while collecting roughly five times more usable training data than traditional teleoperation. According to TechCrunch, the physical hand hardware was produced in partnership with the Chinese firm Wuji Tech, while the system design and data pipeline are Genesis AI's own. The company has also said it plans to reveal a full-body general-purpose robot built on the same technology.
GENE-26.5 is trained through a data engine that fuses three sources: data captured with the dexterous glove, egocentric video from humans wearing cameras, and human-activity videos from the internet. Underpinning the synthetic-data side is Genesis AI's in-house, high-fidelity physics-simulation and rendering stack, which the company describes as ultra-fast and capable of generating training data far faster than real time. This simulation lineage traces back to the open-source Genesis engine, whose published benchmarks claimed simulation speeds up to roughly 430,000 times faster than real-world time. The combination of a custom physics simulator and a custom rendering engine is intended to narrow the sim-to-real gap so that policies learned in simulation work on physical robots.
The Genesis AI company should not be confused with, but is closely connected to, the open-source Genesis physics engine (the GitHub project Genesis-Embodied-AI/genesis-world). Genesis began in December 2024 as an academic research project, the product of a roughly 24-month collaboration that the project's announcement said involved more than 20 research labs (other reports cite contributors from 18 universities). Zhou Xian, later Genesis AI's CEO, was a lead organizer of that collaboration while a PhD student at Carnegie Mellon.
The open-source Genesis is a universal, GPU-accelerated simulation platform written in Python that unifies multiple physics solvers (rigid body, MPM, SPH, FEM, PBD, and others) so it can simulate rigid and articulated bodies, cloth, liquids, smoke, deformables, and more in a single scene. Its authors reported that it ran 10 to 80 times faster than existing GPU-accelerated stacks such as Nvidia's Isaac Gym and MuJoCo's MJX, reached simulation speeds around 430,000 times faster than real time, and could train a transferable robot locomotion policy in about 26 seconds on a single RTX 4090 GPU. The project also described a "generative" layer, in which a vision-language-model agent uses the simulator's APIs as tools to author 4D dynamic worlds, with that generative framework slated for gradual release. Genesis is distributed under the permissive Apache 2.0 license.
The two entities are linked through their shared founder and technical heritage: the open-source Genesis repository states that the project's development is now officially supported by Genesis AI, the company. In practice, the open-source engine is a public simulation platform for the robotics and robot learning community, while Genesis AI is a venture-funded company building proprietary foundation models, hardware, and a commercial data engine that it has said it may partially open-source. The relative-speed and "high-fidelity physics" claims most often attributed to Genesis AI originate from the open-source engine's published numbers.
The name "Genesis" is heavily reused, and Genesis AI (the robotics company) is unrelated to several other things that share the word:
When this article and most technology press refer to "Genesis AI," they mean the robotics startup co-founded by Zhou Xian and Théophile Gervet.
Genesis AI competes in a crowded and well-capitalized field of companies building general-purpose "brains" for robots, an area sometimes called physical AI or embodied AI. Direct comparisons in press coverage include Physical Intelligence, which had raised on the order of $400 million, and Skild AI, another Carnegie Mellon spinout reported at a multibillion-dollar valuation. Genesis AI's emphasis on owning the full stack, from a custom physics simulator and data engine to a dexterous hand and the foundation model itself, distinguishes its approach from peers that focus primarily on the learning model or on humanoid hardware. The broader category also includes simulation-and-data efforts from large incumbents, such as Nvidia Cosmos and the Isaac Lab robot-learning framework, and research programs on large behavior models for manipulation. A Khosla Ventures partner, Kanu Gulati, framed the core bet bluntly at launch, asking whether anyone would build a large robotics foundation model that truly generalizes across tasks.
Robot foundation model · Embodied AI · Physical Intelligence · Skild AI · Nvidia Cosmos · Isaac Lab · Large behavior models · Robot learning · Robot manipulation · Simulation