Every industrial robot today is essentially a student that has been cramming from a textbook written by humans. Engineers collect teleoperation data, label trajectories, and hope the machine generalizes to the messy floor of a real warehouse or home. Veeda AI, a new startup led by former NVIDIA AI research VP Sanja Fidler and two longtime collaborators, wants to move robotics from supervised memorization to interactive learning inside a generated world. The company just raised a $90 million seed round led by Khosla Ventures and Radical Ventures, a figure that signals venture capital is now betting on simulation infrastructure, not just another humanoid hardware team.
What Veeda is actually building
Veeda calls its product a multimodal foundation world model. Unlike the video-generation models that produce pretty clips from a text prompt, Veeda’s model is action-conditioned: it takes a robot’s motor commands or camera motion as input and predicts what the world should look like next. The goal is not a movie, but a trainable simulator where an embodied agent can practice grasping, navigation, assembly, and recovery from failure millions of times without breaking a single real sensor.
The founding trio spent years inside NVIDIA’s spatial-intelligence and generative-AI labs. Fidler ran NVIDIA AI Research and co-led work on 3D scene synthesis; Zan Gojcic specialized in neural reconstruction and high-fidelity simulation environments; Huan Ling worked on real-time high-resolution visual synthesis. Their academic lineage runs through GameGAN, DriveGAN, latent-diffusion video models, and NVIDIA’s Cosmos world-model family. In June 2026 they published OmniDreams, an action-conditioned generative world model for autonomous-driving simulation that could rank driving policies consistently with real-data reconstructions. Veeda is essentially taking that closed-loop philosophy from cars and expanding it to general robots.
Why simulation is the only scalable school for robots
Real-world robot training is a data bottleneck dressed up as an engineering problem. Teleoperation gives high-quality demonstrations but is slow and expensive. Internet videos are abundant but lack action labels. Pure imitation learning works until the robot faces an object in a new pose, a different lighting condition, or an unexpected nudge. At that point the policy stalls or fails silently. Demo-based learning hits the same wall: GEN-1.5 can pick up a new task from a few seconds of demonstration, but every task still requires capturing a real-world demo first.
Veeda’s diagnosis is simple: robots need trial-and-error learning, and trial and error in the real world is too dangerous, slow, and unparallelizable. Drop a robot in a kitchen and it might shatter a plate. Drop a million copies of the same policy inside a neural simulator and they can explore overnight. The company is betting that a sufficiently faithful world model can become the “reality gym” for embodied AI, the same way OpenAI’s Dota environment or DeepMind’s AlphaFold datasets became gyms for other domains.
How the world model is architected
Veeda has not published a full technical white paper yet, but its founders’ prior work and job postings sketch a plausible stack. The core is an action-conditioned video model paired with a latent dynamics model. The generative backbone is expected to use a rectified-flow diffusion transformer, a block-causal autoregressive design, and a causal video tokenizer. The tokenizer’s compression ratio matters a lot: the more it squeezes video tokens, the longer the model can roll out coherent futures without choking on memory.
To make the model useful for robot training, Veeda has to solve four hard sub-problems:
- Action conditioning without teleoperation labels. Veeda plans to use inverse-dynamics and latent-action models to recover pseudo-actions from unlabeled videos, dramatically expanding training data beyond expensive human-operated demonstrations.
- Long-horizon stability. World models drift because they eventually feed on their own outputs. Veeda is expected to use diffusion forcing, self-forcing, and persistent 3D Gaussian or point-map scene representations to keep generated episodes coherent for minutes, not frames.
- Real-time interaction. A simulator that takes minutes to render one step is useless for reinforcement learning. The model would be distilled into a few-step sampler so a policy can act inside the world model at interactive rates.
- Physics-grounded verification. Pure generative models hallucinate. Veeda likely combines physics-based reward models and validators to keep rollouts within physically plausible bounds, following Fidler’s stated belief that the winning path blends traditional physics solvers with learned generation.
This is not a replacement for simulators like Isaac Lab or open-source VLA stacks such as DM0.5. It is an attempt to sit above them: a learned world generator that plugs into existing APIs and can post-train policies such as π0 and GR00T N before they are deployed on real hardware.
From self-driving cars to general robots
The OmniDreams preview matters because it proved the team can already build closed-loop driving simulators where a policy’s action changes the generated scene frame by frame. Moving that capability to robotics means the world model must handle manipulation, deformable objects, contact physics, and diverse camera viewpoints. That is harder than driving, but the economic payoff is larger: warehouses, hospitals, agriculture, and homes all need robots that can adapt without months of re-engineering. With mass production still stuck at the 10,000-unit wall, progress on the training-data layer will decide when that payoff actually arrives.
If Veeda succeeds, the workflow for building a robot policy starts to look like training a large language model. You pre-train on broad simulated experience, fine-tune on a small amount of real data, and evaluate with Sim2Real gap metrics. The expensive part shifts from building bespoke sim scenes to scaling the world model itself.
What this means for the physical AI stack
Veeda’s launch is a signal that physical AI is entering its data-infrastructure phase. For the last few years, the spotlight was on hardware platforms, foundation vision-language-action models, and humanoid demonstrations. Now the frontier is moving toward the layer that generates the training data itself. We have seen a similar pattern before: in autonomous driving, companies that owned realistic simulators could iterate faster than competitors stuck on the road. In gaming, procedural environments enabled AlphaStar and OpenAI Five. Robotics is next.
Investors are effectively underwriting a bet that interactive learning in generated realities will become a core primitive of the embodied-AI stack, alongside actuators, perception models, and inference silicon. If the bet pays off, the next generation of robot foundation models may not be measured only by parameter count, but by how many simulated years of experience they can consume.
Actionable takeaways
- For robotics engineers: Start treating world models as a first-class training environment, not a visualization tool. Expect your Sim2Real pipeline to absorb more compute than the policy itself.
- For AI product leaders: Ask whether your robot’s value comes from hardware novelty or from the data engine that trains its brain. Veeda is evidence that capital is flowing to the latter.
- For investors and policymakers: Watch the fidelity-cost curve of generated simulation. The moment generated environments become cheaper and safer than real-world pilot deployments, regulatory and liability frameworks will have to catch up.
FAQ
What is a foundation world model for robotics?
It is a generative simulator that predicts future sensory observations from current state and robot actions, allowing policies to practice tasks without physical robots.
How is Veeda different from NVIDIA Isaac or traditional physics simulators?
Isaac and similar tools rely on hand-built graphics pipelines and physics solvers. Veeda aims to learn world dynamics from data and combine them with physics constraints, making it easier to scale to new objects and scenes.
When will interactive world-model training become mainstream?
Niche industrial tasks may see production use within two to three years, but general household robots will take longer because the Sim2Real gap remains wide.
Source: Based on reporting by 36Kr / 阿尔法公社.
