NTT Says Its Physical AI Lifts Robot Success Rate to 92.3%, Plans Mujin Baggage Trials
The Japanese telecom group pairs a robot world model with a cheaper way to customize its tsuzumi language model, as its R&D Forum opens on Wednesday.

What to know
- NTT said its new world model raised a small robot's shelf-loading success rate to 92.3% from 73.1% by predicting load collapses and drops before the robot moves.
- NTT plans field trials with robotics company Mujin, starting with loading and unloading airport baggage, but neither NTT nor press reports gave a timetable.
- A separate NTT tuning method lifted tsuzumi 2's accuracy on internal business-manual questions to 83.1% from 65.3%, NTT said on Nov. 12.
TOKYO
NTT said last week it has built a physics-based "world model" that lets robots test several possible moves in simulation before acting, and that the system raised a small robot's success rate at loading irregular items onto a shelf to 92.3% from 73.1%. The Japanese telecom group plans field trials with robotics company Mujin, starting with loading and unloading airport baggage.
The Nov. 11 announcement, followed a day later by a cheaper way to customize NTT's in-house tsuzumi language model, shows where the telecom group wants its AI research money to land: in factories, warehouses and back offices rather than in a head-to-head race with U.S. chatbot makers. Both technologies go on show at the company's invitation-only R&D Forum, which opens on Wednesday at the NTT Musashino R&D Center in Tokyo and runs on Nov. 18 to 20 and Nov. 24 to 25.
How NTT's world model works
Robots that pick and stack objects of different shapes and weights still fail often. Task success rates are often too low for stable operation, Takahiro Matsumoto, a senior researcher at NTT Human Informatics Laboratories, said, according to the trade publication Business Network.
NTT's answer is to let the robot look ahead. Existing control software, either a vision-language-action model or a conventional motion planner, proposes several candidate actions. For each one, NTT's system generates a predicted video of what would happen and scores it, flagging moves likely to make a load collapse or an object fall. The robot then picks the option most likely to work.
The core of the system is what NTT calls a physics control adapter, which estimates each object's position, orientation, velocity and mass and predicts how it will move using equations of motion. That adapter is paired with a video generation model. Because the physics is represented explicitly, NTT said the model can cope with situations and motions that were missing from its training data, which matters in settings where collecting complete training sets is hard.
In the shelf-loading test, in which a small robot stacked irregularly shaped items, NTT combined its world model with the open-source π0.5 vision-language-action model, according to Impress Watch. On OBJ3D, a public benchmark of colliding objects, NTT reported a score of 7.81 on LPIPS, a measure of how closely predicted frames match real ones (lower is better), against 7.92 for a method called OCK, Impress Watch reported. The same report put Nvidia's Cosmos Predict 2.5 at 21.90 on that test.
Those are laboratory results. The shelf test used a small robot, and the benchmark uses synthetic video. Neither NTT's release nor the press coverage gave a start date for the Mujin trials. NTT said it will also explore work with robot service providers and logistics and warehousing operators, and later expand the technology to manufacturing, construction and retail, as a way to ease labor shortages and raise productivity.
A cheaper route to company-specific tsuzumi
The second release, on Nov. 12, targets a different bottleneck: the cost of teaching a language model a client's own procedures. Injecting new knowledge by retraining usually takes tens of gigabytes of text, NTT said, roughly 10 million times the size of a single A4 page.
NTT's method has the model write its own study material. Starting from a small set of business manuals, the model generates training data in varied formats, such as specifications and FAQs, and then learns from them, NTT said. Tested on questions drawn from NTT's internal manuals, tsuzumi 2 scored 65.3% before training, according to the company. Earlier published methods reached up to 73.3%, while the new approach reached 83.1%, NTT said. It added that a 3-billion-parameter version of Meta's Llama 3 model also did better with the new method than with the earlier ones.
NTT plans to roll the technique out across its group companies as one of several ways to tailor its models, the company said. That is the commercial point. In a May update, NTT said tsuzumi 2 can process Japanese business documents with charts and tables on a single graphics processor and is mainly deployed on premises or in private clouds, for customers handling sensitive information. A model that can be fitted to one company's manuals with little data should be easier to sell to those customers.
NTT is not alone in pitching home-grown models to Japanese companies. SoftBank said in November 2025 that its unit SB Intuitions would start offering an API for its 70-billion-parameter Sarashina mini model later that month, aimed at uses such as call-center automation, Business Network reported at the time.
What to watch at the R&D Forum
The forum's theme this year is "Mehr Licht!" (More light!), built around photonic technology for computing, AI and quantum systems. NTT President and Chief Executive Akira Shimada is scheduled to deliver a keynote, titled "Space AIOWN," at 10:30 a.m. on Wednesday in Japanese, with an English version on Thursday. Senior Vice President Shingo Kinoshita follows with a keynote on light and AI.
The forum program also lists a technical seminar on the evolution of tsuzumi on Nov. 24 and interactive demonstrations built on tsuzumi 2. NTT is not streaming the event online; it says recordings of the keynotes will go up on its YouTube channel after the forum.
For investors, the forum will not move NTT's earnings this quarter. The figures to watch are further out: whether NTT names a schedule, sites or customers for the Mujin baggage trials, and how many group companies take up the new tsuzumi tuning method once it is offered to clients.
Sources
- NTT: NTT Develops Physical AI Technology That Predicts the Future Before Robots Move to Prevent Failures (Nov. 11, 2026)
- Impress Watch: NTT, physical AI that predicts the future helps robots avoid failures (Japanese) (Nov. 11, 2026)
- Business Network: NTT develops a world model for physical AI (Japanese) (Nov. 11, 2026)
- NTT: NTT establishes technology to tune LLMs for specialized expertise with small datasets (Japanese) (Nov. 12, 2026)
- NTT: NTT's LLM tsuzumi 2 Updated (May 19, 2026)
- NTT: NTT R&D Forum 2026 (Nov. 17, 2026)
- Business Network: SoftBank and OpenAI joint venture begins full operations (Japanese) (Nov. 5, 2025)
Our standards: Capital Times reports under the principles set out in our Editorial Standards and Ethics Policy. Spotted an error? Report it to our editors.







