NTT World Model Lifts Robot Task Success From 73.1% to 92.3%

NTT, Inc. says a physics-based world model raised robot shelf-loading success from 73.1% to 92.3% in a small-robot test, by screening candidate actions before the machine moves. The company announced the technology on November 11, 2026. The caveat is scale: the result covers a controlled shelf task, not the messy real world of airport baggage.

The first target application is automated loading and unloading of airport baggage. NTT argues that luggage varies too much in shape, weight, and arrangement for conventional robot control alone to predict whether an action will collapse a load or drop an item. Collecting training data for every possible item, the company adds, is impractical. NTT plans field trials with Mujin, Inc., a robotics firm, to test baggage handling directly.

How does the model screen robot actions?

The technology sits alongside existing robot control software rather than replacing it. Candidate actions can come from two sources: a vision-language-action model, which converts camera video and language instructions into robot actions, or a motion planner, which calculates the controls needed to achieve a task objective.

For each candidate, the system generates predicted video and evaluates collisions, drops, and load collapse. It then produces a score, and the control program selects the action judged most likely to succeed. The screening happens entirely before movement begins.

NTT is candid about why it built the system this way. Conventional world models, the company says, do not explicitly represent objects' physical quantities. That gap can produce disappearing or implausibly deforming objects in generated video and inaccurate collision predictions. Their reliance on large training datasets also makes failures more likely in situations absent from that data — precisely the problem baggage handling poses.

NTT's answer is a physics control adapter. It estimates each object's position, orientation, velocity, and mass from video, applies physical laws to calculate subsequent motion, and feeds those calculations into the video generation model. NTT says this explicit representation can handle situations and motions missing from training data. The model also outputs the physical quantities behind each prediction, which the company says makes the basis for a generated result easier to understand.

Takeaway: the system tries to reason from physics, not from memorized examples — a claim the baggage trials must still prove.

What do the benchmarks show?

NTT also measured future-video accuracy on the public OBJ3D benchmark, a dataset of video frames showing synthetic objects colliding. The company reports its model outperformed conventional methods on Learned Perceptual Image Patch Similarity (LPIPS), a metric that compares predicted video against ground truth in a way aligned with human perception. Lower LPIPS values indicate greater similarity.

Both headline results — the 92.3% shelf success rate and the OBJ3D win — come from small-robot tasks and synthetic videos. Neither yet demonstrates performance on real airport luggage with its irregular weights and shapes.

What comes next?

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

The technology is scheduled for exhibition at NTT R&D Forum 2026, an invitation-only event running November 18–20 and November 24–25, 2026. A November 25 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.

The Mujin field trials carry the real weight here. They will test whether a shelf-loading improvement measured on a small robot transfers to airport baggage handling, where physical variation is exactly what motivated the model's development. Until those results arrive, the 92.3% figure is a laboratory number, not an operating record.