Most robots learn from mistakes the expensive way: they act, something falls, and the system tries again. On November 11, 2026, NTT announced a different approach. Its new physical AI technology lets a robot imagine what will happen for several possible moves, then pick the one least likely to end in a collapsed stack or a dropped item.
According to NTT’s press release, the system raised the success rate of a shelf-loading task from 73.1% to 92.3%. The first target is a messy real-world job: loading and unloading airport baggage.
Why robots struggle with unpredictable loads
Physical AI is the name for systems that let machines perceive the physical world through cameras and sensors, then decide and act on their own. A popular building block is the “world model,” a model that predicts how the surroundings will change after an action. NTT says today’s world models have two weaknesses. They do not explicitly represent physical quantities such as mass or velocity, so generated video can show objects vanishing or bending in impossible ways. And they depend on large amounts of training data, so predictions tend to fail in situations the data never covered.
Airport baggage and delivery parcels expose both problems. Shapes, weights and arrangements differ every time, so conventional robot control cannot easily tell in advance whether a move will topple a pile or send a bag to the floor. Collecting training data for every possible item is impractical.
How the physics control adapter works
NTT’s answer is a component it calls a “physics control adapter.” It estimates each object’s position, orientation, velocity and mass from video, then calculates how that object will move next using equations of motion. NTT says it is the first in the world to integrate such an adapter into the video generation model that forms the base of a world model.
The calculated results feed into video generation, which reduces objects disappearing or deforming and improves predictions of collisions and load collapse. The company lists four advantages: predictions grounded in physical laws, the ability to handle situations absent from training data, natural-looking video output that lets people see the predicted result of an action, and explicit physical quantities behind each prediction, which makes the basis for a result easier to understand.
Imagining several futures before moving
The technology is designed to sit beside existing robot control software rather than replace it. A vision-language-action model, a motion planner or another program proposes candidate actions. For each one, the NTT model generates a video of the predicted outcome, scores the likelihood of success, and the robot selects the action most likely to work. NTT says the model is built to run fast enough to fit inside real robot control loops.
That design matters for adoption. A company with a working picking system does not have to rebuild it; it could add a checking layer that vetoes risky moves before they happen.
What the tests showed
NTT ran two evaluations. In a shelf-loading task with a small robot simulating item stacking, applying the technology lifted success from 73.1% to 92.3%. On the public OBJ3D video prediction benchmark, which features collisions among several objects, the NTT model outperformed conventional methods on LPIPS, a perceptual similarity metric where lower values mean closer matches to real video. In scenes where the older model missed a collision, NTT’s succeeded, including a scene not in the training data in which the robot arm blocked the view of objects.
These are results from a small-robot task and a benchmark, not from a live airport. NTT says field trials are still ahead.
What it means
For logistics and warehousing operators, the appeal is reliability. A pick that fails costs time, may damage goods and can force a human to step in. Predicting failure before acting is a cheaper way to improve success rates than collecting ever larger datasets. NTT says it plans to work with Mujin, Inc. on implementation, to consider robot service providers and logistics companies as partners, and later to extend the approach to manufacturing, construction and retail, where labor shortages are a recurring pressure.
For the wider field, the work reflects a growing view that pure data-driven prediction is not enough for robots. Blending learned video models with explicit physics gives systems a better chance at the unfamiliar. The technology is scheduled to be shown at NTT R&D Forum 2026 in November.
Things to watch
- Results from field trials with Mujin on real airport baggage, where weights and shapes vary widely.
- How much computing the model needs when scoring many candidate actions in real time.
- Whether other robot makers adopt the “predict, score, then act” pattern as a layer over existing planners.
Conclusion
NTT’s world model will not make robots perfect, but it points at a practical idea: let the machine rehearse a move in imagination first. If the 92.3% result holds up outside the lab, that rehearsal could become a standard safety net for robots that handle the unpredictable.