Generalist AI Raises $400M to Build Foundation Models That Let Robots Handle Any Task

The robotics industry has a software problem. Despite billions of dollars in hardware investment — humanoid robots, autonomous forklifts, precision assembly arms — most deployed systems are still fundamentally brittle. They work reliably in tightly controlled environments and fall apart the moment a box is placed in the wrong orientation, a surface is wet, or the lighting changes unexpectedly. Solving that brittleness is what Generalist AI was built to do.

On June 4, the San Mateo, California-based startup announced it had raised $400 million in new funding, led by Radical Ventures, at a $2 billion valuation. The round included participation from 8VC, Union Square Ventures, Norwest, Hanabi Capital, Nvidia’s corporate venture arm NVentures, and Bezos Expeditions. New angel investors joining the cap table include AI researcher Fei-Fei Li, Xiaomi co-founder Bin Lin, and entrepreneur Naval Ravikant — a roster that signals the company is attracting significant credibility from both the academic AI and consumer tech communities.

A Team Built From the Best of Robotics AI

The founding team’s credentials are unusual even by the standards of an industry that has produced dozens of well-credentialed startups. Pete Florence, the company’s founder, was a senior research scientist at DeepMind, where he created RT-2 — the robotic transformer model that demonstrated language-conditioned robotic control and became one of the most cited papers in physical AI research. He also co-developed PaLM-E, DeepMind’s early effort to build a large multimodal model specifically for embodied AI applications.

Joining him as Chief Scientist is Andy Zeng, also a DeepMind alumnus, whose research on robot learning with language models helped establish the technical foundations that Generalist AI is now commercializing. Andrew Barry, the CTO, comes from Boston Dynamics, where he worked on the low-level locomotion and manipulation systems that give robots the ability to interact with the physical world reliably.

Together, the trio brings together the three capabilities that physical AI requires: large-scale model training (Florence), robot learning at the system level (Zeng), and the hardware-adjacent engineering that makes software work in a real robot body (Barry).

GEN-1: What the Model Actually Does

Generalist AI launched its first model, GEN-1, in April 2026. The system is designed as a foundation model for robot learning — trained on a combination of robot trajectory data, real-world video, and synthetic environments, with the goal of enabling what the company calls “adaptive intelligence.”

Adaptive intelligence means the robot can observe when a task is not going as expected and adjust its behavior in response. Fold a shirt and the fabric slips — a rigid programmed system fails. GEN-1 allows the robot to observe the slip, reason about the new state of the material, and retry the task with an updated grasp strategy. The same adaptive loop applies to warehouse operations where box dimensions vary, lighting shifts between shifts, and floor layouts change as inventory cycles.

The company claims GEN-1 outperformed Physical Intelligence’s pi-0 model — another well-funded physical AI contender — on task execution speed in comparative evaluations. Physical Intelligence, which raised $400 million of its own last year, is building toward a similar goal of general-purpose robot intelligence, making the competitive comparison a meaningful signal of Generalist AI’s relative standing.

Physical AGI: Bigger Than Humanoids

The framing Generalist AI uses for its market opportunity is deliberately broader than the humanoid robot narrative that has dominated physical AI coverage. The company targets “the layer of artificial intelligence that operates and interacts with the physical world,” a description that encompasses not just humanoid or industrial robots, but autonomous vehicles, drones, smart building systems, and surveillance infrastructure — any system that must convert AI inference into physical action in an uncontrolled environment.

This framing matters strategically. The humanoid robot market, while growing rapidly, is ultimately constrained by the pace of hardware manufacturing and the complexity of bipedal locomotion engineering. By positioning GEN-1 as a cross-platform foundation model that can be licensed or integrated into any physical AI system — regardless of the mechanical form factor — Generalist AI is targeting an addressable market that is several orders of magnitude larger than any single hardware category.

NVIDIA’s participation as an investor, through NVentures, is particularly notable in this context. NVIDIA has been explicit about its ambitions in physical AI through its Cosmos model family and its robotics simulation platform Isaac. An investment in Generalist AI signals that NVIDIA sees value in backing the model layer above its own hardware, creating a potential distribution and co-development relationship as GEN-1 scales.

The Competitive Landscape

The physical AI funding wave of 2025 and 2026 has produced a cluster of well-capitalized startups all pursuing variants of the same thesis: that general-purpose robot intelligence requires foundation model approaches analogous to what transformers did for language. Besides Physical Intelligence, competitors include 1X Technologies (backed by OpenAI and EQT), Covariant (which merged with Amazon-adjacent robotics operations), and Skild AI (another DeepMind spinout targeting general-purpose robot learning).

What distinguishes Generalist AI’s positioning is its explicit focus on the model layer rather than the hardware layer. The company is not building its own robot bodies — it is building the brain that can animate any robot body. That approach reduces capital requirements relative to integrated hardware-software competitors and creates the possibility of rapid customer acquisition across the entire installed base of industrial and collaborative robots worldwide.

The $400 million raise at a $2 billion valuation is a relatively tight multiple compared to some peers in the space, suggesting that Radical Ventures and its co-investors are pricing technical risk appropriately rather than loading up the valuation in anticipation of revenue that has not yet materialized. Generalist AI has not disclosed its current customer or revenue base, though the April 2026 GEN-1 launch has reportedly attracted pilot deployments with several major manufacturing and logistics companies.

If the physical world is the next frontier for AI, Generalist AI’s founding team and funding position have placed the company close to the center of the race to claim it.