Fei-Fei Li's World Labs Bets on Spatial AI for

Language models got their superpowers from the internet — an essentially unlimited ocean of text to train on. Robots don't have that. Every piece of training data for a robot arm means someone has to physically operate it, in real space, at real speed, which is often slower than a human doing the same task by hand. If you want a robot that can sort packages faster than a person, you can't get there by having a person teleop it slowly while you log the data.

That data bottleneck is the central problem Fei-Fei Li and Yunzhu Li sat down to talk about in a recent a16z conversation — and it's the reason World Labs just acquired SceniX.

The acquisition that started with a signup form

The backstory here is quietly delightful. Li (Fei-Fei) founded World Labs two years ago around the idea of spatial intelligence — her term for AI that can generate, reason about, and interact with spaces, whether those spaces are physical or virtual. World Labs has been building a base model called Marble, which takes image or text prompts and outputs geometrically consistent 3D worlds (think: Gaussian splats, meshes, actual renderable space).

Yunzhu Li — no relation, Columbia professor, MIT PhD, former Stanford postdoc under Fei-Fei — founded SceniX to build what he calls a real-to-sim-to-real pipeline for robotics. The idea: scan a real environment, reconstruct it digitally with tight physical alignment, train and evaluate robots inside that digital world, then deploy back to reality.

The twist? SceniX didn't come to World Labs as a merger target. They came as a customer.

"When we released the first version of Marble last winter," Fei-Fei explained, "SceniX just signed up — as a customer. And I didn't even know what it was. And then I realized this is Yunzhu's company." The synergy became obvious fast enough that what started as a vendor relationship turned into an acquisition.

It's the kind of origin story that sounds clean in hindsight. Worth noting, though: a16z has investments in this space, and this conversation happened on their platform. That context doesn't invalidate anything said — but it's worth having in your head as you read the enthusiasm.

What "real-to-sim-to-real" actually means

The pipeline Yunzhu describes goes like this: you capture a real environment — its appearance, geometry, and crucially its dynamics (how things change when you push them) — and you reconstruct it digitally with enough fidelity that behavior in the sim reliably predicts behavior in the real world. You train your robot policies there, evaluate them there, and then you ship to real hardware.

The key word he keeps returning to is alignment. Not the AI safety kind — the more literal kind: does what happens in the simulation actually correspond to what would happen if you deployed the robot for real? If your sim says a checkpoint achieves 95% task success and the real robot achieves 60%, your sim is useless. If the numbers track, your sim is enormously powerful.

Yunzhu frames the value in two buckets: reliability and efficiency. Reliability comes from systematic coverage — in simulation you can vary lighting, friction, object geometry, and physical parameters at will, exposing your model to edge cases that would be expensive or dangerous to recreate in the physical world. Efficiency comes from the fact that simulation isn't bound by gravity or robot speed limits. Clients who need robots moving faster than human pace can train policies at simulated speeds that real teleoperation can never produce.

For evaluation specifically — "eval" in AI parlance means understanding how well a model performs at a specific checkpoint, like does it succeed 90% of the time or 99%? — the gap is stark. Robotics evals in the real world require atoms to move through space, which means they're "multiple orders of magnitude slower" than iterating on a language model. Every robotics demo video playing at 8x speed is kind of a tell for this problem.

The "sim-to-real gap" objection

The obvious pushback — and host Martin Casado raised it directly, citing Sergey Levine — is that simulation always eventually diverges from reality. Sim-to-real transfer is a real challenge in robotics research. If your digital world is wrong in ways you can't anticipate, your robot learns the wrong thing.

Yunzhu's response is that the framing is a false binary. "They don't contradict with each other," he said. Simulation and real-world data aren't competing approaches — they're complements at different stages of a data flywheel. Early in training, you lean on physics-heavy simulation for structure and consistency. As you accumulate real deployment data, you shift toward more learning-based models that can capture nuance physics engines miss. The mix evolves over time.

Fei-Fei adds the human cognition angle: "Think about human intelligence. We do a lot of simulation in our head. There's a very important role simulation plays that real world data doesn't play — which is counterfactual reasoning." You play out scenarios that haven't happened and learn from them. Every soccer coach diagramming plays on a whiteboard is running a sim. Waymo, she notes, runs billions of simulation hours — and is more simulation-heavy than real-data-heavy.

The "sim doesn't have to be perfect" point is also worth sitting with. Yunzhu points to locomotion robots — quadrupeds walking on snow or through bushes — as evidence. Nobody simulated every individual snowflake. The simulation captured enough of the structure to enable transfer. The question is always about the right level of fidelity, not perfection.

What this response doesn't fully resolve is whether the manipulation tasks SceniX is targeting (warehouse picking, electronics assembly) have the same forgiving transfer properties as locomotion. Locomotion is periodic and relatively low-dimensional. Manipulating deformable objects or unknown items is messier. That's an open empirical question, and honest researchers will tell you it's still being worked out.

The humanoid hype correction

One of the more grounded moments in the conversation is Yunzhu's taxonomy of robot environments: fully structured (car factories, fully controlled) → semi-structured (Amazon warehouses, restaurants, electronics assembly lines) → unstructured (your home, his home, especially with three dogs and a five-year-old).

His argument: humanoids are optimized for unstructured environments because that's what human bodies evolved for. But unstructured environments are the hardest problem in robotics — maximum variation, minimum controllability. Betting on humanoids for near-term commercial deployment means betting on solving the hardest version of the problem first. That's a rough order of operations.

Fei-Fei frames it cleanly: "From a business point of view, from a pragmatic technology point of view, the unstructured environment and a generalized body is actually the hardest problem to solve. It's not necessarily even the right way to solve the problem."

This is a meaningful methodological stake in the ground. Semi-structured environments — warehouses, assembly lines, restaurants — have enough predictability to make simulation alignment tractable, and enough scale to build a real business on. SceniX's platform is designed to be embodiment-agnostic (it works with arms, mobile manipulators, grippers, whatever) and model-agnostic (it can train from scratch or fine-tune existing vision-language-action models). The bet is on being infrastructure, not a specific robot.

What World Labs is actually building

It's worth being precise about this, because "World Labs acquires robotics company" conjures images of robot arms coming off assembly lines. That's not what's happening. "That's not building a robot," Casado noted. "It's building an environment which another company can place their robot brain in to navigate and to learn."

Exactly right. The product is the world — the digital reconstruction, the simulation infrastructure, the training and evaluation substrate. The clients bring their own robots, their own models, their own end-use cases. World Labs and SceniX provide the environment those systems learn and get tested in.

The foundation model question is genuinely open. Fei-Fei said World Labs is building a base model, and the natural evolution of that is a multimodal model that can take frames, depth, text, and actions as both input and output — essentially a forward simulator (given this action, what happens next?) plus a policy model (given this goal, what action should I take?). That's the vision. It's also not shipped yet, and "we're not ruling this out" is doing some work in that sentence.

Yunzhu's north star is simpler and more grounding: "My north star is to make robots work. In the real environment." When they surveyed the public about what they wanted robots to do, a third of the responses were about cleaning. People want the dirty jobs handled. That's the mundane, unsexy, actually-achievable version of robotics progress — and it's where SceniX has been spending its time.

The "right measured optimism," as Fei-Fei put it, is hard to calibrate from the outside. The technical case for simulation-driven robotics training is real and has precedent in self-driving. The open questions — sim-to-real fidelity for manipulation tasks, whether the data flywheel closes fast enough to compete with pure real-world data approaches, whether semi-structured is a stepping stone or a ceiling — are exactly the ones that will determine whether this thesis pays off.

Two years from now, World Labs says success looks like a small number of validated customers in specific verticals proving the infrastructure actually works. That's a concrete enough bar to hold them to.

Yuki Okonkwo is the AI & Machine Learning correspondent at Buzzrag. She covers the people building the algorithms and the systems they're setting loose on the rest of us.