# NTT reports 92.3% success in robot shelf-loading test; field trials remain ahead

> NTT says a physics-informed model raised success in a small-robot shelf-loading simulation from 73.1% to 92.3%. Field trials with Mujin remain planned, and the release does not report a commercial deployment.

- Canonical URL: https://newsverum.com/en/news/ntt-robot-physical-ai-shelf-loading-test-2026-11-11
- Publisher: Newsverum
- Section: Technology
- Published: 2026-11-12T03:09:09+00:00
- Updated: 2026-11-12T03:09:09+00:00
- Event / data period: 2026-11-11
- Other editions: [Türkçe](https://newsverum.com/tr/news/ntt-robot-physical-ai-shelf-loading-test-2026-11-11.md), [العربية](https://newsverum.com/ar/news/ntt-robot-physical-ai-shelf-loading-test-2026-11-11.md), [Français](https://newsverum.com/fr/news/ntt-robot-physical-ai-shelf-loading-test-2026-11-11.md), [Español](https://newsverum.com/es/news/ntt-robot-physical-ai-shelf-loading-test-2026-11-11.md), [Deutsch](https://newsverum.com/de/news/ntt-robot-physical-ai-shelf-loading-test-2026-11-11.md), [Português](https://newsverum.com/pt/news/ntt-robot-physical-ai-shelf-loading-test-2026-11-11.md)

NTT said on 11 November that a physics-informed world model improved the success rate of a shelf-loading task performed by a small robot from 73.1 percent to 92.3 percent. The task simulated stacking items, and NTT described the result as a comparison with and without its technology. The company also reported a lower LPIPS score than two reference systems on the OBJ3D video benchmark. Its first planned field use is airport baggage handling, but field trials with robotics company Mujin are still ahead. The release does not describe a commercial deployment.

The system is designed to review candidate robot actions before movement. NTT says its model estimates object position, orientation, velocity and mass from video, then uses equations of motion to forecast how objects may move or interact. A proprietary physics control adapter connects those calculations to a video generation model. For each action proposed by existing control software, the system generates a possible future scene and scores outcomes such as a collision, a dropped object or a collapsing stack. It supplements an existing controller rather than establishing that robots can safely operate without supervision.

The shelf-loading result is notable, but the absolute gain is more informative than a broad claim that the robot became better at work. The reported success rate rose by 19.2 percentage points, calculated as 92.3 minus 73.1. Relative to the starting rate, that is about 26.3 percent, calculated as 19.2 divided by 73.1. Both calculations use rounded values in NTT's release. NTT does not state how many trials were run, how success was defined, whether the items or starting conditions varied, or how much results varied across runs. The release also provides no uncertainty interval.

NTT's Table 1 gives more detail on a separate prediction benchmark. It lists an LPIPS score multiplied by 100 of 7.81 for the NTT model, 7.92 for OCK, an academic model, and 21.90 for NVIDIA Cosmos Predict 2.5, a commercially available video generation model. Lower LPIPS is better. The NTT value is 0.11 lower than the OCK figure, or about 1.4 percent relative to 7.92. That arithmetic describes the displayed metric only. The release supplies no statistical interval or protocol comparison that would establish whether the small gap is robust across settings.

The benchmark itself needs context. A 2025 ICCV paper by Yeon-Ji Song and co-authors describes OBJ3D as a synthetic dataset with rigid objects, neutral grey scenes, 2,920 training sequences and 200 test sequences. It was designed to examine object geometry and interactions such as collisions. NTT says it evaluated its model on the public OBJ3D benchmark, but does not say whether it used the same data split, test protocol or training setup as the OCK paper. This result therefore measures video prediction in a controlled benchmark, not baggage handling in a real airport.

The technical distinction NTT emphasizes is explicit physical state. A video model can produce visually plausible frames while missing the forces or motion that determine whether a load will tip. NTT says its adapter represents mass and motion and applies equations of motion to generate candidate futures. That approach could help a robot reject a risky plan before acting, especially when each box or suitcase differs. The press release does not provide a latency figure, hardware configuration, memory demand, or safety test, so it cannot establish whether these calculations are fast and reliable enough for an operating warehouse.

NTT calls luggage and parcel handling an intended application and says it plans field trials with Mujin. It also says the technology is scheduled for display at NTT R&D Forum 2026. These are development plans, not evidence that airport operators have selected the system, installed it or measured labour savings. The release gives no customer, project budget, installation date, throughput target, error rate in live work, or price. The step from a simulated stacking task to mixed baggage at a belt or cargo area would require tests under changing lighting, occlusion, conveyor movement and human presence.

The announcement is best read as a research milestone with a concrete but narrow test result. The robot task suggests a possible operational use, while OBJ3D offers a separate check of predicted video quality. Neither result alone establishes general reliability across logistics, manufacturing or retail, the sectors NTT lists for possible expansion. Before deployment, decision-makers would need repeated trials, transparent definitions of task success, safety controls, performance across unseen objects and a comparison of the full system with current alternatives. NTT has described a route toward those trials, not their outcome.

## Data

### Shelf-loading simulation success rates

Rates NTT reports for its simulated stacking task with a small robot.

Methodology: NTT reports task success rates without and with its technology; it does not disclose the number of trials, success criterion or variability. 92.3 minus 73.1 equals 19.2 percentage points. 19.2 divided by 73.1 is about a 26.3 percent relative increase. These calculations use rounded company values and are not separate measurements.

| label | value | unit | period | basis | value_qualifier |
|---|---|---|---|---|---|
| Small-robot shelf-loading success rate without NTT technology | 73.1 | percent | 2026-11-11 | observed | exact |
| Small-robot shelf-loading success rate with NTT technology | 92.3 | percent | 2026-11-11 | observed | exact |

Sources: <https://group.ntt/en/newsrelease/2026/11/11/261111b.html> · Download data: [JSON](https://newsverum.com/api/v1/articles/ntt-robot-physical-ai-shelf-loading-test-2026-11-11/data/shelf-loading-task-success-rate?lang=en&format=json) · [CSV](https://newsverum.com/api/v1/articles/ntt-robot-physical-ai-shelf-loading-test-2026-11-11/data/shelf-loading-task-success-rate?lang=en&format=csv)

### OBJ3D video benchmark LPIPS comparison

Model scores shown in NTT’s published table. Lower LPIPS is better.

Methodology: NTT’s Table 1 reports LPIPS multiplied by 100, with lower scores better. Score cells preserve the published values 7.81, 7.92 and 21.90 with the scale label. The release gives no uncertainty intervals or direct protocol comparison between NTT and OCK.

| Model | Model type | Reported LPIPS x100 score |
|---|---|---|
| NTT world model | Company model | 7.81 LPIPS x100 |
| OCK | Academic model cited by NTT | 7.92 LPIPS x100 |
| NVIDIA Cosmos Predict 2.5 | Commercial video generation model | 21.90 LPIPS x100 |

Sources: <https://group.ntt/en/newsrelease/2026/11/11/261111b.html>, <https://openaccess.thecvf.com/content/ICCV2025/papers/Song_OCK_Unsupervised_Dynamic_Video_Prediction_with_Object-Centric_Kinematics_ICCV_2025_paper.pdf> · Download data: [JSON](https://newsverum.com/api/v1/articles/ntt-robot-physical-ai-shelf-loading-test-2026-11-11/data/obj3d-video-benchmark-lpips?lang=en&format=json) · [CSV](https://newsverum.com/api/v1/articles/ntt-robot-physical-ai-shelf-loading-test-2026-11-11/data/obj3d-video-benchmark-lpips?lang=en&format=csv)

## Sources

- [NTT, NTT Develops Physical AI Technology That Predicts the Future Before Robots Move to Prevent Failures](<https://group.ntt/en/newsrelease/2026/11/11/261111b.html>) (Source checked: 2026-11-12; Published: 2026-11-11)
- [Computer Vision Foundation, OCK: Unsupervised Dynamic Video Prediction with Object-Centric Kinematics, ICCV 2025 paper](<https://openaccess.thecvf.com/content/ICCV2025/papers/Song_OCK_Unsupervised_Dynamic_Video_Prediction_with_Object-Centric_Kinematics_ICCV_2025_paper.pdf>) (Source checked: 2026-11-12)
- [Computer Vision Foundation, OCK supplementary material](<https://openaccess.thecvf.com/content/ICCV2025/supplemental/Song_OCK_Unsupervised_Dynamic_ICCV_2025_supplemental.pdf>) (Source checked: 2026-11-12)

---

Cite the canonical URL above. [Editorial standards and corrections](https://newsverum.com/en/editorial)
