Veeda Raises $90M to Build Robot Training World Models
The former Nvidia researchers aim to create simulated worlds where robots can learn safely through large-scale trial and error.
A virtual training ground for physical AI
Veeda AI has emerged from stealth with more than $90 million in seed financing to develop world models for robotics. The company is led by Sanja Fidler, formerly Nvidia’s vice president of AI research, alongside longtime collaborators Zan Gojcic and Huan Ling. Khosla Ventures and Radical Ventures are backing the startup.
Veeda’s proposed product is not a general-purpose robot or a conventional simulation package. It wants to build generative, multimodal models that can produce physically plausible virtual environments from sensor and real-world data. Robots would receive virtual bodies inside those environments, allowing their control systems to interact, make mistakes and improve without damaging equipment or endangering people.
The company argues that imitation learning from recorded human behavior will not be sufficient for broadly capable robots. Its alternative is interactive learning at scale: exposing machines to many variations of an environment and letting them test actions repeatedly. Veeda describes this simulated-reality layer as shared infrastructure that could support robot training and evaluation across manufacturing, logistics, transportation and other physical industries.
Fidler’s team brings experience spanning generative AI, computer vision, 3D reconstruction and robotics simulation. Veeda lists operations in Toronto, Mountain View, Zürich and Singapore, but has not yet disclosed a model, dataset, customer or release schedule. The financing therefore represents a substantial bet on the team and technical thesis rather than a demonstrated commercial system.
Why it matters
Robotics is increasingly constrained by data rather than hardware alone. Real-world experiments are slow, costly and difficult to reproduce, while hand-built simulators struggle to match the variety and visual complexity of reality. If generative world models can provide sufficiently accurate physics, perception and interaction, they could become for robotics what large-scale training corpora became for language models. Veeda’s unusually large seed round makes it one of the best-funded new efforts to build that foundational layer, although its central claims remain to be validated in working systems.