NTT DOCOMO develops ‘Physical AI coordination platform’ linking AI agents and robots

NTT DOCOMO announced Oct. 27, 2026, that it had developed a “Physical AI coordination platform” linking personal agents with multiple robots. It uses A2A, a communications protocol for AI agents, to manage the process from receiving instructions to a robot’s actions. A demo using the platform is scheduled to be shown at NTT R&D FORUM 2026, which begins Nov. 18, 2026.

NTT DOCOMO said the platform keeps users’ attribute information within the system and sends robots only action instructions that are permitted. Its design incorporates attribute-based controls and action-history tracking into the connection between agents and robots. This relates to the challenge of managing whose instructions led to which actions in scenarios involving rules for individual users, such as the restaurant food-delivery example. However, the use cases presented are future concepts, not results from testing in real-world environments.

The key point of the platform is that it does more than connect robots to AI agents: It also places checks on whether actions are allowed based on user attributes and generates behavior suited to the surroundings within the same process. It uses the existing A2A protocol for connections, while DOCOMO developed the permission checks, history tracking and behavior-generation functions. In other words, the system combines standardized connectivity with functions for handling site-specific rules and controls. The adoption of A2A alone does not confirm operational compatibility among different robots, and the announcement did not specify supported models or connection requirements. Permission controls and consideration of the surroundings are separate processes. The former compares attribute information with rules for a given use case and blocks unauthorized actions. The latter generates behavior based on surrounding conditions. DOCOMO described a design in which attribute information is not passed to the robot and only action instructions are sent, but it did not say what information the behavior-generation process uses or how generated actions are verified. History recording is positioned as a way to trace instructions and actions afterward, but the announcement did not disclose how long records are kept or who can access them. The restaurant food-delivery example combines service controls based on age and allergy information with actions to avoid people. However, this is a future use case, not a record of deployment or safety testing in an actual restaurant. The platform appears to function less as a source of the robot’s own capabilities and more as a management layer that checks agent instructions against rules, passes them to robots and links them to behavior suited to the situation. In assessing practical use, key questions will include who sets the rules, how the system selects among multiple robots, the accuracy of its recognition of surrounding conditions and how it handles conflicts between permission decisions and generated actions. A related article from this newspaper, NTT DOCOMO develops Physical AI platform that determines actions based on user attributes and surrounding conditions, was also published Oct. 27, 2026, and described decision-making based on attributes and surrounding conditions. The explanation of the platform in this announcement confirms the use of A2A connections, in-platform management of attribute information and action-history tracking. The related issue is shifting from the introduction of decision-making functions itself to whether rules can be applied and action histories tracked consistently when different robots are connected. The NTT R&D FORUM 2026 exhibition will provide an opportunity to demonstrate the system, but DOCOMO said it plans to assess its practicality through internal deployment, joint research and demonstration experiments. The scope and performance of real-world operation remain to be established.

At the core of the platform, the “A2A Gateway” checks user attribute information, such as age, against rules registered for each use case and blocks unauthorized actions before they are carried out. Attribute information is managed within the platform, and only action instructions are sent to robots. The platform also records whose instructions prompted an action, when it was carried out and what action was taken, and generates behavior based on surrounding conditions. It uses A2A for connections, while DOCOMO developed the permission-control, history-tracking and behavior-generation functions. The exhibition will be held Nov. 18-20 and Nov. 24-25, 2026, at the Event Hall on B1 of the NTT Musashino R&D Center. The exhibit number is C10. Examples of future uses include selecting a delivery robot based on a restaurant order and controlling food delivery according to user attributes.