# NTT DOCOMO's Physical AI Gateway: A2A Robot Permissions and Demo Plans

DOCOMO has developed a platform connecting personal agents with robots, using permission checks and action histories. A public demonstration is planned for November.

By Paras Madan | 2026-10-28


NTT DOCOMO announced a Physical AI platform on October 27 that connects personal agents with robots and checks requested actions before execution. It is at the demonstration stage, giving builders a design to examine rather than a generally available service.

The interesting part for agent builders is the decision between understanding a request and allowing a machine to act on it. A valid instruction can still be inappropriate for the person making it or the place where it would be executed.

## How DOCOMO's A2A Gateway works

The [official announcement](https://www.docomo.ne.jp/info/news_release/2026/10/27_02.html) describes using Agent2Agent, or A2A, for communication, with DOCOMO-developed permission controls and action tracking.

The gateway compares user attributes with setting-specific rules, forwarding only permitted instructions. Attributes remain within the platform, according to DOCOMO.

It records who requested each action and when, and generates behavior intended to account for the surroundings. These are company-described capabilities, not independently measured safety outcomes.

## What A2A provides, and what needs its own checks

The [A2A discovery documentation](https://a2a-protocol.org/latest/topics/agent-discovery/) explains how agents advertise capabilities through Agent Cards and find services they can collaborate with. It also recommends protecting sensitive card information with authentication and authorization, and using dynamic credentials rather than putting static secrets into cards.

That provides a common communication structure. The decision about whether a particular physical action is allowed still needs rules for the actual environment. Knowing that a robot offers a delivery capability does not establish that every user can request every delivery.

An illustrative restaurant workflow makes the distinction clear. Before an agent sends a robot to a table, the system may need to check who placed the order and whether the requested action is permitted. Those checks need trustworthy attributes and current rules. If either is wrong, a correctly implemented gateway can still reach the wrong decision.

## Demonstration timing and availability

DOCOMO plans a demonstration at NTT R&D FORUM beginning November 18, followed by further validation. Public SDK access, API pricing and a general launch date are not specified.

That leaves the announcement at a development-and-demonstration stage. It offers a design to examine, but not enough evidence to assume compatibility with a particular robot fleet or a production operating environment.

## The questions a deployment must answer

Permission checks are most useful when they remain enforceable through the full action path. A deployment evaluation should establish what happens when attributes are missing, a rule service is unavailable or a robot cannot complete an authorized action safely. It should also show how an operator stops execution and how the system records the outcome.

Those are evaluation questions, not claims that DOCOMO's platform fails them. The announcement's contribution is to make permission enforcement and traceability explicit parts of connecting software agents to machines. The November demonstration can show how that design behaves beyond a diagram, while production readiness will require evidence from defined tasks and operating conditions.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders explain agent infrastructure and robotics updates in content that makes the design, access and limits clear.

[Book a 30-minute call](https://cal.com/paras-madan/30min).