Generating 3D worlds is no longer enough; we must make robots act within them.

The startup World Labs acquires SceniX to link its 3D models to robotic simulation. A step toward closed-loop physical training for machines.

Known for generating explorable 3D worlds with Marble, Fei-Fei Li’s spatial intelligence startup, World Labs, is acquiring SceniX, a young robotics simulation company. The deal marks its biggest step yet toward embodied AI—the kind that acts in the physical world rather than on a screen.

SceniX, founded by Columbia researchers Yunzhu Li and Changxi Zheng along with Sonny Hu, builds high-fidelity simulations where a robot learns a task in a virtual environment before executing it on real hardware. Its tools are used to generate training data and evaluate robots, including on edge cases. Its starting premise is simple: a robot is relatively easy to model, but the environment in which it operates is much harder, and that is where a successful transition from simulation to reality is decided.

The complementarity is clear. World Labs knows how to build persistent 3D worlds, while SceniX approaches the same territory through robotics and simulation. The stated bet: the next breakthroughs will come from combining World Models, learning-based simulation, and a closed loop with the real world, where what the robot learns there feeds back into the simulation. Fei-Fei Li summarizes the shift as moving from generating worlds to interacting with them.

World Labs, which closed a $1 billion round in February, is thus shifting its trajectory from 3D world creation toward physical AI. Technical details of what SceniX has already built are promised soon.