A “Matrix” for robots: Sanja Fidler co-founds Veeda AI

Veeda AI wants to train robots inside world model simulations, multiplying trial-and-error runs without relying on real-world hardware.

Teaching a robot to perform a task in the real world gives every mistake a physical cost. A flawed trajectory can damage a component, interrupt a production line, or put someone at risk. Experience is also difficult to accelerate: a robotic arm can only test a limited number of actions during one hour of real time.

Veeda AI wants to move this learning stage into simulated environments generated by world models. The company was co-founded by Sanja Fidler, Zan Gojcic, and Huan Ling, three researchers who worked together at NVIDIA on 3D vision, scene reconstruction, simulation, and generative models for Physical AI.

Their starting point is that embodied intelligence will not advance through imitation alone. Human demonstrations can teach a machine to reproduce actions found in its training data, but they struggle to cover every mistake, variation, and unfamiliar situation it may encounter. Veeda wants robots to act, observe the consequences, fail, and try again.

This loop already exists in several forms of robotic learning. The stated difference lies in scaling it. Within a simulation, multiple versions of the same robot can test different strategies simultaneously, accelerate time, and encounter rare scenarios without wearing down physical equipment. A task performed once in a warehouse can be replayed with different objects, positions, lighting conditions, obstacles, and behaviors.

Traditional simulation platforms already allow robots to train and undergo evaluation inside virtual environments. These systems often rely on 3D scenes, physical properties, and scenarios created in advance. Their diversity therefore depends on the time required to build each environment and program its variations.

Veeda is positioning itself at this creation layer. Its future multimodal world models are intended to learn from large volumes of sensor and physical-world data, then generate environments varied and coherent enough to host embodied agents. The company describes virtually unlimited worlds in which robotic intelligence could explore different ways to complete a task.

A world model built for robotics cannot simply produce a visually convincing scene. It must preserve object states, represent distances, contacts, and motion, then anticipate the consequences of an action. If a robot pushes a box, picks up a tool, or collides with an obstacle, the environment must evolve consistently and return observations compatible with the robot’s sensors.

The ambition covers both training and evaluation. The same simulation could expose several robotic models to comparable situations, measure their ability to recover from mistakes, and identify edge cases before real-world deployment. Target applications include logistics, manufacturing, construction, and the transportation of people and goods.

The founders’ backgrounds place Veeda in direct continuity with NVIDIA’s world model research. Sanja Fidler led the company’s AI research and Spatial Intelligence Lab in Toronto. She is also an associate professor at the University of Toronto and a founding member of the Vector Institute.

Zan Gojcic led research focused on reconstructing complete environments from sensor data, with a background in 3D vision and point-cloud processing. Huan Ling managed teams working on generative world models and high-fidelity foundation models. He also contributed to NVIDIA Cosmos, a platform developed for simulation, autonomous vehicles, and robotics.

This research pedigree is the main concrete information currently available. Veeda has not yet released a model, demonstration, benchmark, or technical documentation. Its training data, environmental representation, and future access options also remain undisclosed.

Moving from simulation to the real world remains a central challenge. A virtual scene can appear coherent while reproducing contacts, friction, material resistance, or sensor behavior imperfectly. The more a robot exploits an approximation specific to the simulation, the more its performance may deteriorate when it encounters real machinery, flooring, or objects.

Veeda is backed by Khosla Ventures and Radical Ventures, although no funding amount has been disclosed. The company is hiring across four locations: Toronto, Zurich, Mountain View, and Singapore. For now, its project remains a research infrastructure under construction, designed to give robots a place to make millions of mistakes before the first one happens in the physical world.