NTT’s Physics-Based World Model Screens Robot Actions Before Execution

On November 11, 2026, NTT, Inc. announced a physics-based world model that evaluates candidate robot actions before execution to reduce failures such as dropped objects and collapsing loads, saying the technology raised task success in a small-robot shelf-loading test from 73.1% to 92.3%.

The first envisioned application is automated loading and unloading of airport baggage. Variation in luggage shape, weight, and arrangement makes it difficult for conventional robot control alone to determine in advance whether an action will collapse a load or drop an item, and collecting training data for every possible item is impractical. NTT plans field trials with Mujin, Inc. for baggage handling.

How the Model Evaluates Candidate Actions

The technology works with existing robot control software, adding prediction and evaluation before movement. Candidates can come from a vision-language-action model, which converts camera video and language instructions into robot actions, or from a motion planner, which calculates the controls needed to achieve a task objective. For each option, it generates predicted video and evaluates collisions, drops, and load collapse, producing a score that lets the control program choose an action assessed as likely to succeed.

NTT says conventional world models do not explicitly represent objects’ physical quantities, which can produce disappearing or implausibly deforming objects in generated video and inaccurate collision predictions. Their dependence on large training datasets also makes failures more likely in situations absent from that data.

NTT’s physics control adapter estimates properties including each object’s position, orientation, velocity, and mass from video, applies physical laws to calculate subsequent motion, and supplies those calculations to the video generation model. NTT says this explicit representation can address situations and motions absent from training data. The model also outputs the physical quantities behind a prediction, making the basis for the generated result easier to understand.

Benchmark Results and Expansion Plans

NTT also evaluated future-video accuracy on the public OBJ3D benchmark, which contains video frames of synthetic objects colliding. NTT reports that its model outperformed conventional methods on Learned Perceptual Image Patch Similarity (LPIPS), a metric comparing predicted video with ground truth in a manner aligned with human perception; lower values indicate greater similarity.

Beyond baggage and parcel handling, NTT plans to pursue logistics, warehousing, manufacturing, construction, and retail applications and will consider working with robot service providers and logistics and warehousing companies. Reduced labor requirements and improved productivity are objectives for that expansion, not outcomes established by the reported test.

The technology is scheduled for exhibition at NTT R&D Forum 2026, an invitation-only event running November 18–20, 2026, and November 24–25, 2026. A November 25, 2026, seminar on AI robotics and the physics-informed world model lists Mujin Japan CEO Isamu Arase and NTT executive research engineer Jun Shimamura as speakers.

NTT’s announced results cover a small-robot task and synthetic collision videos. The planned Mujin trials will test whether the reported shelf-loading improvement transfers to airport baggage handling, where pre-execution screening must address the physical variation that motivated the model’s development.