NTT gives robots a physics world model that previews a move before they make it, lifting a lab stacking test from 73.1% to 92.3%

The backstory: a robot can watch a lot of video and still get the physics wrong

Once a robot has cameras and an AI model, it can look at a scene and decide how to reach and place an object. Many of these systems use a "world model": a preview of what the scene will look like if the robot takes a given action, before it actually moves.

The catch is that a lot of world models mainly guess the next video frames from huge training sets, without explicitly tracking physical quantities such as an object's mass, orientation, or speed. The preview can then do odd things: objects vanish, shapes warp, or a collision never happens. Unfamiliar boxes and stacks — the ones missing from the training data — are where these systems often fail. Airport bags and delivery parcels are exactly that kind of scene. Shapes, weights, and arrangements change constantly, and no one can film every case for training.

What NTT says it built

On November 11, NTT announced a world model aimed at physical AI, a shorthand for machines that sense the real world with cameras and sensors, then decide and act on their own.

At the center is what NTT calls a physics control adapter. It estimates each object's position, orientation, velocity, and mass from video, computes how the object should move next using equations of motion, and feeds those results into a video-generation model that produces a clip of what would happen if the robot took that action. NTT says it is the first to integrate this kind of adapter into the video model at the core of a world model. That is the company's own claim, not an independent ranking.

The system is also meant to sit on top of software that already exists. NTT says it can work with current robot controllers, vision-language-action models, or motion planners: those programs propose candidate actions, the world model generates a future video for each one, scores risks such as a drop, a collision, or a collapsing stack, and selects a move that is more likely to succeed. The company also says the model outputs the physical quantities behind a prediction, which makes it easier to inspect why one action looked safer than another.

The test results, and their limits

NTT's headline experiment used a small robot on a shelf-loading task meant to simulate stacking. With the new model, the success rate rose from 73.1% to 92.3%. The company also says the model beat conventional methods on LPIPS, a similarity score designed to track human perception, on the public OBJ3D benchmark, which includes collisions among multiple objects. The press release does not list the specific LPIPS numbers.

Those figures come from NTT's own lab test and benchmark comparison, not from an independent trial on an airport floor. The stacking test used a small robot in a simulated loading task. That is not the same as a system already working steadily beside a baggage carousel.

What comes next is still a plan

The first use NTT has in mind is loading and unloading airport baggage of mixed shapes and weights. It plans to move into field trials with the robotics company Mujin, and says the same approach could later spread to logistics, warehousing, manufacturing, construction, and retail. The technology is scheduled to be shown at the NTT R&D Forum in November 2026.

What it could mean for you

If "preview, then move" holds up around real bags and parcels, warehouse and airport robots might drop fewer items and stop less often. For travelers, more careful automated handling could mean luggage is stacked less roughly. For logistics companies short on workers, it would be one more tool to try.

That point has not arrived yet. What is confirmed today is that NTT improved a simulated stacking test in the lab and intends to take the system into the field. Real baggage is messier, heavier, and less uniform. The field trials are the next result to watch.

Original source published: NTT, Nov 11, 2026, 14:00:30 Beijing time (UTC+8)