World Labs CEO Fei-Fei Li and SceniX co-founder Yunzhu Li used a new a16z interview to lay out the logic behind World Labs’ acquisition of SceniX, casting robotics as the first serious test case for what Li calls “spatial intelligence.”
In the interview, Li said World Labs was founded two years ago as a startup and frontier model lab with a clear goal: make spatial intelligence the next frontier in AI. In her description, that means giving AI the ability to generate, understand, reason and interact in real or virtual spaces.
World Labs sees large world models as one path toward that goal. Robotics, in turn, is where those ideas meet the physical world.
Why robotics sits at the center of the plan
Li said acting in space is not limited to robots. In fields such as VFX, gaming and design, people already build and interact with immersive virtual worlds. World Labs has long held the view that the world people inhabit can become a multiverse of spaces where users, creators and developers can act.
Still, she argued that the ability to act in physical space remains one of the most exciting and consequential capabilities for future AI systems, and robots are central to that effort. That is why World Labs has treated robotics as the first proving ground for spatial intelligence and as an important application for world modeling.
Bringing in SceniX, she said, fits directly into that long-term mission.
SceniX is focused on the data bottleneck
Yunzhu Li, who is also an assistant professor at Columbia University, said his work has been shaped by a simple objective: help robots perceive the physical world better and interact with it in ways that let them work in real environments.
He said the biggest bottlenecks for general-purpose robotics today sit in training and evaluation. SceniX is building a real-to-sim-to-real pipeline that maps real environments into digital worlds aligned closely enough that events in simulation correspond to what can happen in the physical setting.
That alignment matters because it allows at least part of the data collection and evaluation workload to move from the real world into scalable digital environments.
According to Yunzhu Li, SceniX has built a team spanning robotics, robot learning, simulation and rendering, with the aim of assembling a complete real-to-sim-to-real stack.
Fei-Fei Li added that the relationship between the two companies did not begin as an acquisition discussion. SceniX first came to World Labs as a customer. After World Labs released the first version of its generative model Marble last winter, around November or December, SceniX signed up and began using it. Li said she did not initially realize the company was Yunzhu Li’s, and only later called him after seeing the overlap between the two groups.
Where Marble and SceniX meet
Li described Marble as a foundation model that World Labs continues to train and iterate. The current public version can take one or more images, text, or a combination of those modalities and convert them into a geometrically consistent world.
She contrasted robotics with language models. Large language models can draw on abundant internet-scale data. Robotics cannot. In her view, SceniX is addressing one of the sector’s hardest problems: the lack of training and evaluation data. If robotics is going to benefit from scaling laws, the industry has to answer a more basic question first: where that data will come from.
Li said the two teams are highly complementary. She named SceniX co-founder Changxi Zheng, a Columbia professor with deep simulation expertise and a VFX background, and Sonny Hu, an engineering leader who worked at a startup later acquired by Amazon and then on several computer vision stacks at Amazon.
Her assessment was straightforward. Yunzhu Li brings robotics leadership and full-stack capability from hardware to systems. Zheng brings simulation strength. World Labs contributes generative modeling and computer-vision-based 3D reconstruction, areas SceniX needs. The combined footprint is broader than either side had on its own.
Yunzhu Li gave a more technical version of the same point. SceniX’s real-to-sim-to-real work requires dense reconstruction of environments, including appearance, geometry and dynamics, meaning how a scene changes after an action is applied. That remains expensive and heavy. World Labs, he said, has deep capability in sparse reconstruction and generation, creating room to use Marble and related tools to make environment reconstruction and modeling more efficient.
A robot foundation model is not off the table
Asked whether World Labs may eventually release a robot foundation model, Fei-Fei Li said the company is building foundation models and is not ruling out that direction.
Yunzhu Li said a robot foundation model is, by nature, a multimodal model. It needs to handle frames, text, images, depth and other modalities, with action as a particularly important one.
If frames and actions are both used as inputs, he said, the model can act as a simulator and predict how an environment will change after a specific action. If action is treated as the output, the same setup becomes a policy model, one that predicts what actions a robot should take in the real world to move toward a specific goal.
He described this as an “Omni-Model” that could help embodied AI researchers and developers understand environment modeling and decision planning, while also serving as a base model for downstream fine-tuning in specific robotics applications where customers care about reliability and efficiency.
Why they are leaning into 3D and simulation
Many robotics companies are working with video-first approaches. Yunzhu Li argued that 3D and simulation matter because robots need worlds with consistency across space, time, viewpoints and interactions.
That is where he sees strong synergy with Marble. A generated world, if it is internally consistent, can provide part of the infrastructure robots need.
He gave a simple example: if a robot tries to push an object forward and the object suddenly disappears, the model is not providing a usable training signal. Video prediction models are improving quickly, he said, but SceniX wants to build infrastructure that can kick-start a data flywheel. The system can begin with more simulation-driven models, move toward robot policy models that work in real environments, collect new data and feed that data back into the loop.
He also rejected a strict divide between physics-based approaches and learning-based ones. A model, in his framing, can sit in between. It can capture the essential structure of a problem and keep improving as more data arrives.
When asked about his own “north star,” Yunzhu Li said he is pragmatic and wants robots that work in real environments. He recalled building a benchmark with Fei-Fei Li during his postdoctoral period and asking the public what they wanted robots to do. Out of more than 1,000 tasks collected, about one-third involved cleaning. Those are jobs people do not want, and they remain a target category for robotics.
Fei-Fei Li said she admires how practical the SceniX founders are. Though most of them came from academia, they moved early to work with designers and customers in real industry settings, including labs, warehouses and electronics assembly environments.
Models do not need to be perfect
On the question of precision, Yunzhu Li said neither the environment model nor the robot model has to be perfect. What matters is whether the model captures the essential structure needed for transfer to the real world.
He pointed to drones, robot vacuums, quadrupeds and humanoids as examples of systems that rely on models and on simulation-to-real transfer. A robot that walks through snow or brush does not need a simulator that reproduces every twig and every patch of snow exactly. It needs one that captures the structure that matters and supports the right kinds of randomization.
SceniX, he said, is studying the level of fidelity needed to model the world around a robot so that training in simulation and digital environments can transfer back into physical settings.
Simulation and real-world data are not opposing choices
Some researchers argue that simulation will always drift from the physical world, making real-world data indispensable. Yunzhu Li said the two are not in conflict. Simulation is a way of predicting how the environment changes after an action, which is another way of saying it is a model of the world.
That model does not have to be purely physical. It can combine physics and learning. He said SceniX is collecting real-world data as well and expects to use those data at different stages of a data flywheel.
At the beginning, the work may lean more heavily on physics to help robots learn the structure of the world and the right learning rhythm. As more data comes in through collection efforts and customer work, environment modeling can shift toward a more learning-driven setup that combines the strengths of physics, geometry and consistency with the power of data and compute.
Fei-Fei Li argued that simulation is not an either-or question. She compared it to how humans think before acting. People run many simulations in their heads, and one of simulation’s key functions is counterfactual reasoning: learning from events that have not happened, cannot happen directly, or do not have enough real-world data attached to them.
She said robots can benefit from the same principle. As an example, she pointed to Waymo, which she said has stated that it used billions of hours of simulation data and built a heavily simulation-dependent path rather than relying only on real-world driving data. If simulation matters there, she argued, it plainly has an important role in robot learning more broadly.
Yunzhu Li broke the case for simulation into two gains: reliability and efficiency.
- On reliability, robot systems need data that covers the range of states and changes they may encounter. Simulation allows systematic randomization and control over lighting, geometry types and physical parameters, which helps cover that space.
- On efficiency, many teams still collect data through teleoperation. With current teleoperation devices and exoskeleton setups, data collection can even be slower than a human simply doing the task directly. Customers want robots to move faster than humans, but speed is not solved by just increasing motor commands because gravity stays constant and dynamics do not disappear. In simulation, robot behavior can be accelerated systematically while still training against changes in environmental dynamics.
His conclusion was that simulation brings distinct value on both fronts.
Training and evaluation sit at the center of the platform
Asked how SceniX uses its platform, Yunzhu Li said there are two core functions: training and evaluation.
Evaluation, he said, is often overlooked in robotics. But if a team is training a robot model, it has to know how that model is performing, and evaluation is what drives iteration.
He made the point with a checkpoint example. If a team is trying to distinguish between a checkpoint that performs at 90% and another at 92%, the important question is how long it takes to verify the difference. If the team has to deploy every model in the physical world to test that 2-point gap, the cycle becomes painfully slow.
Real-world robot evaluation today moves orders of magnitude more slowly than language model evaluation, he said. Robotics tasks are also diverse by nature. Robots must act, move in space and obey physics. He noted that many robot demo videos are shown at 8x or 10x speed because the systems themselves are slow.
Real-world evaluation is not just slow. It is also dangerous and expensive. That is why some customers want digital environments for robot evaluation. Since SceniX’s digital environments have shown alignment with the real world, better performance in simulation is likely to signal better performance in real settings as well. That creates a path to safer, faster and more scalable evaluation.
On the training side, controllability is central. The company wants to control state, parameters, lighting, friction and other physical variables, as well as changes in object geometry. Training has to cover enough diverse scenarios to generate informative data and build robustness, and that level of control is hard to achieve systematically in the physical world.
Teleoperated data collection is slow and constrained by the number of robots, the number of remote-control devices and the operational burden around data itself. Simulation environments, by contrast, are fully controllable and analyzable. In a digital world, developers can define the coverage of a data distribution and know whether a robot is reliable inside that range. Yunzhu Li said that certainty, efficiency and scalability explain why customers want to train robot systems in digital environments.
Fei-Fei Li said World Labs had already seen signs of that demand before SceniX approached the company. Marble had attracted inbound interest from robotics companies, including both early model-building teams and companies aiming at practical deployment, even though World Labs was not yet in a position to serve that market directly.
The infrastructure is meant to be model-agnostic and robot-agnostic
Yunzhu Li said the company is building infrastructure and software that let developers create worlds where robots can learn and be evaluated. The stack is not tied to a specific model class or a specific robot body.
That means single-arm or dual-arm systems, fixed-base or mobile-base robots, and different grippers or complex end effectors should all be able to run in the same digital environments. The goal is simple: help them operate reliably and efficiently in real-world settings.
He applied the same reasoning to model choice. The data generated by the platform can train new models from scratch or fine-tune popular VLA, or vision-language-action, models and World Action Models. His view was blunt: the specific model matters less because models will keep changing, while infrastructure will last.
In that sense, SceniX is not choosing sides on hardware and is not making a single algorithmic bet. It is trying to build a general digital world where many kinds of robots can gain capabilities that transfer to real deployment.
Deployment is likely to begin in semi-structured settings
On commercialization, Yunzhu Li said real-world robotics usually follows a progression from fully structured environments to semi-structured ones and then to unstructured environments.
Factories and automotive production lines are fully structured settings where variables are known and controlled, and those settings have already been automated for decades. Semi-structured environments include places such as Amazon warehouses, restaurants and hotels, where tasks can be adapted for robots and the environment is broadly controlled, but unexpected items still appear. Unstructured environments, such as homes, remain the field’s “holy grail.”
Robustness comes from sufficient coverage of scenarios, he said. At least for now, tackling semi-structured environments first is more feasible than going straight into fully unstructured ones. The long-term destination may be the latter, but the near-term route is more pragmatic.
Fei-Fei Li approached the question from morphology. Humanoid robots mirror the human form because humans evolved for unstructured environments. But human fingers and legs are not the best tools for every single task. Humans ended up with a highly general body plan, not a task-optimized one, because that generality was useful for survival in unstructured settings.
Commercially and technically, she said, a general body operating in an unstructured environment is one of the hardest combinations to solve. For some narrow problems, a more specialized body may be the better answer. That leaves SceniX with the challenge of staying neutral across robot forms while making infrastructure that works across different semi-structured environments.
Human-level energy efficiency is still far away
Asked whether robots can reach human-like energy economics in everyday physical labor anytime soon, Yunzhu Li said it will take a long time. Making robots work in real environments is ultimately a systems problem involving hardware, software and the “brain,” down to details such as finger friction coefficients. Getting all of that to work requires many components to come together and continued iteration.
He said what keeps him optimistic is the speed at which robotics learning has moved. The problems he works on now are very different from the ones he studied during his PhD years, which he sees as evidence that the broader ecosystem is evolving quickly and that the pieces needed to build capable robot systems are beginning to converge.
But he said expectations still need to be calibrated. Progress will continue, though human-level overall capability and energy efficiency remain farther out.
Fei-Fei Li made a similar point from the AI side. One of the hardest things in AI right now, she said, is maintaining optimism that is properly calibrated. Even large language models do not match the energy efficiency of the human brain, which she put at roughly 30W.
Integration will be gradual, with customer proof points as the near-term goal
Yunzhu Li said joining World Labs changed SceniX’s trajectory in a deep way. Working together should make the full environment-modeling process more efficient and more scalable.
He contrasted robotics with current language models. Even if language models can produce impressive results, many users would still hesitate to let one book a flight or hotel without reviewing the output. Robotics is different. A robot model that is meant to work out of the box has to be reliable in the real world from the start.
Right now, he said, the field lacks both the data and the complete infrastructure needed for that level of reliability. Building scalable digital worlds where robots can learn and be evaluated is what could unlock that next stage.
Fei-Fei Li said integration will be deliberate rather than immediate. SceniX already has a well-developed technical stack, its own customers and products, so World Labs is not rushing to merge every codebase and team overnight. But integration will happen. The two sides are already discussing simulation, possible foundation models and action-conditioned models, and SceniX is using Marble as an internal customer.
She also said Yunzhu Li will move to San Francisco with part of the team. Once that happens, World Labs will become a company that spans both U.S. coasts. The headquarters is in San Francisco, and the company plans to open an office in New York, which Li said should help with East Coast recruiting. The company has also discussed placing robots in both offices to test and refine a remote robot operation stack that it expects customers will eventually need.
As for the two-year target, Li was specific. She said she would be very happy if World Labs and the SceniX team can secure validated customers in a small number of important vertical use cases and show that their systems and infrastructure genuinely help those customers meet automation needs. Those users, in her words, would become lighthouse cases for broader expansion.
Reference links mentioned in the source material
Full interview:
https://www.youtube.com/watch?v=-tabaM5l3s0
Reference links:
- https://marble.worldlabs.ai/
- https://www.worldlabs.ai/blog/scenix
The source material states that the article was adapted from an edited transcript of the interview published by the WeChat account Quantum Bit and republished by MarsBit.