NTT Docomo develops Physical AI platform that determines actions based on user attributes and surrounding conditions

docomo.ne.jp reported on October 27, 2026, that NTT Docomo had developed a Physical AI platform that determines actions based on user attributes and surrounding conditions. We have not been able to verify the primary source, so please refer to the original report for details.

In its announcement, Docomo described a system that uses user attributes and rules for different situations to block unauthorized actions before they are carried out, while controlling tasks such as food delivery without passing attribute information to the robot. When multiple robots are used in settings such as restaurants, the system’s operational distinction is that it determines in advance who may have what action performed for them, beyond simply connecting the robots, and lets operators track execution history.

The platform’s key feature is that it combines in one system not only connections between AI and robots, but also permission checks before actions are carried out, records afterward and consideration for the surroundings during operation. A2A is an existing protocol used for connections; Docomo developed the permission controls, action-history tracking and behavior generation. The layer that accommodates differences in how robots connect should therefore be viewed separately from the layer that applies rules for each user and situation. Our related article, NTT Docomo develops Physical AI integration platform connecting personal agents and robots, also covers plans to connect personal agents with multiple robots and enable safe, comfortable operation in the same spaces as people. The structure detailed in the latest official announcement keeps attribute information within the platform and sends only action instructions to robots. Where information is accessed and what information is passed to the robot are design considerations that distinguish this system from a simple connection platform. In the restaurant example, the platform handles both restrictions on food delivery based on rules related to age and allergies, and evasion when a person approaches. The former is a control that determines before an action whether an instruction is appropriate; the latter generates behavior in response to the surroundings. They have different purposes and use different information for their decisions. Combined with activity logging, the design makes it possible to track instructions through execution, but the announcement does not specify access permissions for records, retention periods, or procedures for registering and updating rules. A2A compatibility alone does not show how consistently the design will work across robots from multiple manufacturers. The announcement says that developing individual connections had been a challenge, but does not identify supported models or the scope of connection tests, explain how behavior generation will be evaluated, describe how incorrect judgments will be handled, or set out the criteria to be tested in demonstrations. How the accuracy of rule application and consideration for the surroundings are evaluated through internal deployment, joint research and demonstration experiments will be a key factor in assessing the system’s practicality.