NTT's physics-based world model lifts a robot's shelf-stacking success from 73% to 92%, and Mujin will test it on airport baggage

NTT has built a world model that predicts whether a load will collapse or an object will fall before a robot moves, and says it raised success on a small robot's shelf-stacking task from 73.1% to 92.3%. The Tokyo company announced the work on Nov. 11. It plans field trials with Mujin, starting with loading and unloading airport baggage of different shapes and weights.

The design bets that physics, not more video, keeps a world model honest about boxes. Conventional world models do not track objects' physical quantities, NTT says, so in their generated video objects can vanish or deform, collisions go wrong, and scenes outside the training data break them. NTT's "physics control adapter" estimates each object's position, orientation, velocity and mass from video, works out the next motion from equations of motion, and feeds the result into a video-generation model. For baggage, where loads differ in shape and weight, collecting training data that covers every item is not realistic, the company says. It is the opposite wager to NVIDIA's GR00T 2, which learns mostly from human video.

It checks the planner's moves instead of replacing the planner

The model does not drive the robot. A VLA model or a motion planner proposes several candidate actions. NTT's model generates video of what each would lead to, scores it for drops and collisions, and the robot takes the one most likely to succeed. NTT says it is built to run fast enough to sit inside the control flow, and because it outputs the physical quantities behind each prediction, operators can see why it expects a fall.

That is what makes Mujin a natural first partner. Mujin's controllers already plan mixed-case palletizing for industrial arms, and a checker can sit on top of that software without replacing it. On OBJ3D, a public video-prediction benchmark built around multi-object collisions, the model beat earlier methods on the LPIPS similarity measure, NTT says.

The baggage trials are where the 19-point gain has to survive. A small robot stacking items on a shelf is a long way from a full-size arm loading suitcases whose weight the model must guess from video alone. NTT will show the system at its R&D Forum 2026, Nov. 18 to 25.