World Labs and SceniX are building digital worlds to train and evaluate robots

This a16z interview brings together Martin Casado, World Labs co-founder and CEO Fei-Fei Li, and SceniX co-founder Yunzhu Li after World Labs acquired SceniX. It is about combining generative world models with a real-to-sim-to-real robotics pipeline so robots can learn and be evaluated inside controllable digital environments. It is important because robotics does not have the abundant internet-scale training data that accelerated language models, and reliable simulation could provide the scale, coverage, and measurement needed to move robot policies into real deployments.

World Labs calls its field spatial intelligence: AI that can perceive, generate, reason about, and interact with physical or virtual spaces. SceniX approaches the same problem from the robotics side. Its aim is practical—build environments in which robot policies can train, fail, improve, and prove that they work before the corresponding hardware is trusted in the physical world.

World Labs had already identified robotics as a central application of spatial intelligence. Li says acting in physical space is one of the most important capabilities for future AI, but World Labs had concentrated on the generative-model and 3D-reconstruction side of the stack.

Its first product, Marble, turns text, images, video, or rough 3D layouts into persistent, navigable 3D worlds. In the interview, Li describes the public model as generating geometrically consistent environments that can be represented as Gaussian splats or meshes.

SceniX adds the capabilities World Labs did not yet have: full-stack robotics, learning-based simulation, robot-policy training, evaluation, and direct work with industrial customers. Yunzhu Li is an assistant professor at Columbia University whose work focuses on helping robots perceive and interact with the physical world. SceniX's other technical founders bring simulation, rendering, engineering, and deployment experience.

The relationship began through product use rather than a planned acquisition. SceniX signed up for Marble as a customer. World Labs then saw that its sparse reconstruction and generative world models could complement SceniX's denser reconstruction of real environments. World Labs supplies the world-generation layer; SceniX supplies the robotics system that can turn those worlds into training and evaluation infrastructure.

World Labs' formal acquisition announcement makes the same combination explicit: spatial intelligence, world models, learning-based simulation, and a closed loop with real-world learning.

SceniX's real-to-sim-to-real pipeline begins with a physical environment. It reconstructs that environment as a digital world, including its appearance, geometry, and dynamics. A robot policy can then train or be evaluated in the simulation before the resulting behavior is tested on real hardware.

Alignment between the digital and physical environments is the key requirement. SceniX does not need a perfect copy of every leaf, patch of snow, or fabric fold. It needs a model that preserves the structure relevant to the task: what can move, how actions change objects, where collisions occur, and which variations alter the robot's success.

Marble could make the reconstruction step faster and more scalable. SceniX's dense reconstruction is comparatively heavy. World Labs can generate or complete spatially coherent environments from sparse inputs, allowing more worlds and more variations to enter the training loop.

This also explains why the interview treats a consistent 3D world as different from a sequence of plausible video frames. If a robot pushes an object, that object cannot disappear or change position inconsistently when the viewpoint changes. The world must remain coherent across space, time, viewpoints, and interactions because the robot is using it to learn cause and effect.

World Labs says it is building a foundation model, but the speakers do not announce a finished general-purpose robotics model. They describe the ingredients such a model would require.

A robotics foundation model must be multimodal. It may take frames, text, images, depth, robot state, and actions as inputs. When an action is an input, the model can act as a forward simulator, predicting how the environment will change. When an action is the output, the model can act as a policy, choosing what the robot should do to move closer to a goal.

The proposed stack does not require a choice between hand-built physics and learned models. Early systems can use geometry, contact physics, and explicit simulation to preserve the task's essential structure. As deployments collect more data, learned dynamics can become a larger part of the model. The intended loop is simulation to policy, policy to real-world execution, new data back into the model, then another iteration.

Nor is World Labs planning to manufacture one robot. SceniX describes its infrastructure as embodiment-agnostic and model-agnostic. Customers may use fixed arms, mobile manipulators, grippers, quadrupeds, bipeds, or other hardware. They may train a model from scratch or adapt an existing vision-language-action model. World Labs wants to supply the worlds and the software around them.

Li argues that simulation and real-world data are complementary. Real experience reveals what actually happened. Simulation can test what has not happened, cannot safely happen, or is too rare and expensive to collect repeatedly.

Yunzhu Li divides simulation's value into reliability and efficiency. Reliability comes from systematic coverage. A simulator can vary lighting, friction, geometry, object types, physical parameters, and task conditions, then measure whether a policy continues to work. This does not prove the robot will succeed everywhere, but it makes the tested distribution explicit and exposes failure modes that a small real-world dataset may miss.

Efficiency comes from running more experiments without waiting for physical hardware. Teleoperation data is expensive and often collected no faster than a person can perform the task. Real robot evaluation is slow, dangerous, and difficult to reproduce. In simulation, many environments and policies can be exercised in parallel, behavior can be accelerated while preserving relevant dynamics, and checkpoints can be compared against the same controlled conditions.

The interview therefore treats evaluation as an equal bottleneck to training. A robotics team needs to know whether one checkpoint succeeds 90% of the time and another succeeds 90.1%, across enough scenarios to make that difference credible. Repeating the physical task enough times may take too long. Digital environments can make the comparison faster and safer, provided performance in simulation remains aligned with performance in the corresponding real setting.

World Labs has already demonstrated parts of this workflow outside the acquisition. Its published robotics case studies show Marble-generated environments being integrated with Isaac Sim, MuJoCo, and RoboSuite for locomotion, manipulation, training, and repeatable evaluation.

The speakers reject the idea that general-purpose humanoids will move directly into arbitrary homes. Yunzhu Li describes robotics as a progression from structured environments, to semi-structured environments, and finally to unstructured environments.

Factories are structured because operators control most relevant configurations. Warehouses, restaurants, and hotels are semi-structured: the environment can be designed around the task, but the robot must still handle meaningful variation. Homes are unstructured. Their objects, layouts, people, pets, clothing, and interruptions create a much larger space of possible situations.

Humanoid bodies add another difficulty. The human form is general enough to survive unstructured environments but is not optimized for one narrow task. A specialized machine can solve a constrained industrial job more efficiently. SceniX's near-term role is therefore not to bet on one embodiment, but to make its infrastructure work across different robots and semi-structured environments.

Yunzhu Li expects substantial progress but says human-level robotic efficiency and capability will take longer. A useful robot is a whole system whose hardware, software, policy, sensors, contacts, and physical details must work together reliably. Language-model output can still be reviewed before use; a deployed robot has to act correctly in the moment.

World Labs also plans to integrate SceniX deliberately. The teams will connect simulation, generative models, and action-conditioned models, while SceniX continues serving existing customers rather than immediately merging every codebase and product.

Fei-Fei Li defines success over the next two years in commercial terms: validated customers in a small number of important verticals, with the combined infrastructure proving useful to their automation needs and becoming lighthouse examples for a larger business. The customers SceniX is targeting are already close to deployment and have practical tasks whose automation would create immediate value.

The acquisition is therefore both ambitious and constrained. World Labs is extending spatial intelligence from generating worlds to building worlds that can train machines. Its first test is not whether it can produce a universal household robot, but whether its digital worlds can make specific robots learn and prove themselves faster.