HONOR Robot Phone Opens Reservations and Brings Embodied AI to Smartphones

Introduction

HONOR used the 2026 World Artificial Intelligence Conference in Shanghai to show a more complete version of its Robot Phone, a smartphone that combines an articulated camera system with a mobile AI agent.

The company also announced two new color options and opened reservations through its sales channels in China. The Chinese names can be translated as Moon Shadow Gray and Star Trail Silver, although official English retail names may differ when the commercial product reaches additional markets.

HONOR describes Robot Phone as the world’s first robot smartphone and as an early hardware carrier for its new Agentic OS architecture. Those descriptions are the company’s own positioning. The final commercial device, complete specifications, supported regions, price, and feature availability still need to be confirmed through the official product launch.

What is already clear is the direction: HONOR wants the smartphone to move beyond responding through a screen. With a compact motorized camera, multimodal perception, cross-app task execution, and expressive physical movement, Robot Phone is designed to connect software intelligence with visible action in the physical world.

Image credit: HONOR and GSMA/MWC.

Robot Phone Moves From Concept to Reservation

HONOR first presented Robot Phone publicly during MWC 2026. At that stage, the device was introduced as a new smartphone form factor built around embodied AI, robotic motion, and cinematic imaging.

At WAIC 2026, the company shifted the story from an experimental hardware concept toward a product entering the reservation stage.

The device centers on a camera module mounted to a compact robotic gimbal. When deployed, the camera can move independently of the phone body. This allows it to stabilize video, follow a subject, maintain framing during a video call, and use motion as a visible form of interaction.

HONOR’s official product page describes the hardware as an industry-smallest four-degrees-of-freedom gimbal system powered by a highly compact motor. Three axes are used for mechanical camera stabilization, while the additional movement capability supports embodied interaction.

The “industry’s smallest” designation comes from HONOR’s laboratory comparison, not from an independent certification. The official page also notes that the demonstrated functions are still being prepared and that the final interface and user experience may change before commercial release.

From a Smartphone Camera to an Embodied AI Interface

A conventional smartphone can see through its cameras, hear through its microphones, reason through local or cloud models, and trigger digital services.

Robot Phone adds another layer: physical motion.

The camera assembly can turn, tilt, follow a person, and produce gestures that resemble body language. HONOR says it can nod when agreeing, shake its head when disagreeing, move to music, and maintain visual attention during calls.

This does not make the phone a general-purpose robot. Its physical reach is limited to the camera and gimbal mechanism. Still, that mechanism gives the AI system a visible actuator—something it can move in response to what it perceives.

The result sits between several familiar product categories:

Conventional Smartphone	Gimbal Camera	Robot Phone
Primarily screen-based interaction	Mechanically stabilizes footage	Combines stabilization with AI-directed motion
Fixed camera orientation	Camera follows controlled movement	Camera can track subjects and express responses
AI mainly acts through apps	Hardware focuses on imaging	AI connects digital tasks with physical behavior
User holds or positions device	Gimbal improves framing	Device can maintain attention more independently

The practical value will depend on whether the movement feels useful rather than decorative. Reliable tracking, low noise, durability, battery impact, drop protection, and safe retraction are likely to matter as much as the novelty of the mechanism.

Agentic OS: An Operating System Organized Around Intent

The hardware demonstration was part of a larger announcement about Agentic OS, which HONOR calls a partner-oriented multimodal agent operating system.

Traditional mobile operating systems are organized around applications. A user decides which app to open, navigates its interface, enters information, switches to another app, and manually carries context across the workflow.

HONOR’s proposed Agentic OS model begins with the user’s intended outcome.

For example, rather than opening several applications to organize a birthday evening, the user could describe the goal. The system would then need to understand the request, collect preferences, find suitable services, coordinate reservations, handle missing availability, and return a result for confirmation.

HONOR has described four foundational characteristics for Agentic OS:

1. Intent-centric architecture: The system works toward a task rather than treating each app as an isolated destination.
2. Natural interaction: Voice, vision, gestures, gaze, movement, and other sensory inputs can become part of the interface.
3. Proactive intelligence: Agents can plan steps, recommend actions, and execute approved operations.
4. Native cross-device coordination: Phones, PCs, tablets, wearables, home devices, cloud models, and specialized agents can work together.

At the WAIC forum, HONOR expanded this into a full-stack operating-system concept. Hardware, kernels, models, agent frameworks, interfaces, and ecosystem services are treated as connected layers.

In that design, an operating system does not manage only processes, threads, memory, and conventional applications. It also needs to manage models, agent permissions, task state, tools, context, memory, and recovery.

Image from the original Sina Finance article, credited there to Wenhui Daily.

Cross-App Tasks Demonstrated at WAIC 2026

HONOR demonstrated Robot Phone through both a conference forum and hands-on exhibition experiences.

One scenario combined three common birthday-planning tasks:

1. Order a cake.
2. Arrange transportation to a restaurant.
3. Reserve a private room at a karaoke venue.

The point of the demonstration was not that any one action is technically difficult. Existing apps already support each service.

The challenge is maintaining context across them.

A capable mobile agent needs to remember the occasion, location, time, party size, user preferences, budget, and previous choices. It then needs to interact with several services without losing the original goal.

HONOR also showed conversational interaction around football topics and a music scenario in which the camera gimbal moved to the rhythm.

These demonstrations were presented as examples of perception, reasoning, execution, and feedback working together. They do not yet establish how reliably the system handles unusual interfaces, payment confirmation, service failures, network interruptions, or ambiguous instructions.

Why Long-Horizon Mobile Tasks Are Difficult

The article used a travel-planning request to explain the complexity of a real mobile agent:

Arrange a trip to attend a smart-manufacturing conference in Shanghai next Wednesday.

A human understands that this request contains many hidden decisions. A phone agent must make those decisions explicit and connect them to tools.

A complete workflow may require the system to:

1. Identify the correct date and conference location.
2. Retrieve the user’s transport preferences.
3. Search for a suitable train.
4. Prefer a window-seat second-class ticket with a shorter journey.
5. Find a hotel close to the venue.
6. Recover if the preferred train is sold out.
7. Interpret a photo of the conference schedule.
8. Detect and open a QR code.
9. Complete a registration form.
10. Present consequential actions for user approval.

This is a long-horizon task because the later steps depend on earlier results. A sold-out train changes the arrival time, which may change the hotel choice and registration schedule.

Failures are not limited to the language model misunderstanding the request. The task can also break because an app changes its interface, a page loads slowly, a skill returns incomplete data, a login expires, or an action requires confirmation.

A useful mobile agent therefore needs more than a strong model. It needs robust orchestration.

Model and Harness Must Evolve Together

Alibaba and HONOR described the mobile agent as a combination of the underlying model and a surrounding agent harness.

The model is responsible for capabilities such as:

• Understanding the user’s intention.
• Interpreting images and text.
• Planning a sequence of actions.
• Comparing alternatives.
• Making decisions under changing conditions.

The harness manages the operational side:

• Context.
• Long-term and short-term memory.
• Skills and tools.
• Application state.
• Task progress.
• Permissions.
• Error handling.
• Retries and recovery.
• Execution traces.

A model can produce an excellent plan and still fail if the harness cannot operate an interface reliably. Conversely, a well-engineered harness cannot repair every misunderstanding from a weak model.

The harder technical problem is making both parts improve through feedback from real tasks.

When the environment returns a result, the system needs to understand whether the action succeeded, failed, or produced an unexpected state. That information must then influence the next decision.

This feedback loop is what turns a sequence of tool calls into an agent capable of completing work.

Alibaba and HONOR Build a Qwen-Based Mobile Solution

HONOR announced that it is working with Alibaba on a mobile model solution optimized around the Qwen model family.

The collaboration is intended to adapt model capability to real smartphone scenarios rather than simply placing a general cloud chatbot inside a mobile interface.

HONOR’s official WAIC announcement says the companies are jointly developing a terminal-model solution for phone use cases. The Sina/Wenhui report says the resulting work has been applied to Robot Phone.

Alibaba Cloud has separately stated that its work with HONOR improved visual question-answering accuracy by nearly 40%. That percentage is an Alibaba-reported result. The company has not publicly supplied enough benchmark detail in the cited announcement to independently reproduce the figure.

Visual question answering matters for a mobile agent because much of the phone environment is visual. The system may need to read an app screen, understand a photograph, interpret a document, identify a button, inspect a QR code, or combine visual content with stored preferences.

The collaboration also reflects a broader industry pattern. Device makers control sensors, operating systems, accounts, and hardware. Model providers contribute language, multimodal reasoning, training infrastructure, and agent engineering. Neither side can deliver a reliable mobile agent alone.

A Benchmark Based on More Than 1,000 Mobile Tasks

Alibaba said it has designed a benchmark around more than 1,000 real mobile scenarios.

According to the source report, the evaluation covers:

• More than 200 commonly used mobile tools.
• More than 200 memory-retrieval tasks.
• More than 40 combinations of everyday skills.
• More than 400 collaborative interface operations.

The companies said the jointly optimized solution improved task-completion rates and user experience compared with the base model.

No complete benchmark paper, dataset, scoring methodology, baseline table, or independently audited result was linked in the source article. The figures should therefore be described as reported evaluation coverage rather than a public benchmark that external researchers can currently reproduce.

Even so, the choice of evaluation categories is useful. A credible mobile-agent benchmark should test more than whether a model can identify the next button.

It should include:

• Multi-step completion.
• Memory accuracy.
• Cross-app context transfer.
• Recovery from changed or unavailable options.
• Permission handling.
• Confirmation before consequential actions.
• Visual understanding.
• Response to interface changes.
• Time and resource cost.
• User correction and takeover.

A high score on a controlled screen-navigation dataset does not automatically translate into reliable everyday performance.

OpenClaw Turns the Gimbal Into an Agent Actuator

HONOR says Robot Phone’s gimbal is connected to the OpenClaw open platform.

The idea is to expose the camera mechanism as an actuator that developers can control through natural-language-driven agent workflows.

This could allow an application to request actions such as:

• Turn toward a speaker.
• Keep a moving person in frame.
• Reposition for a video call.
• Perform a predefined expressive movement.
• Follow an object while recording.
• Combine perception with a physical response.

The exact public API, developer requirements, permissions, supported commands, and release status were not detailed in the source article.

Physical controls require stricter safeguards than ordinary content generation. A developer platform should define movement limits, collision handling, camera privacy indicators, user consent, emergency stops, foreground/background permissions, and restrictions on silent activation.

Because the actuator includes a camera, the system must also make it obvious when sensing or recording is active.

A Four-Degree-of-Freedom Gimbal Inside a Phone

Robot Phone’s core mechanical feature is the compact 4DoF gimbal.

HONOR says the phone uses a self-developed micro motor whose size has been reduced by 70% compared with a mainstream industry reference selected by HONOR Lab. The company presents it as the smallest micro motor and smallest 4DoF smartphone gimbal in the industry.

The product page separates two related functions:

• Robot-grade embodied motion control, which enables expressive and interactive movement.
• Three-axis gimbal stabilization, which keeps the camera stable during recording.

HONOR also advertises:

• AI object tracking.
• Rapid reacquisition after temporary obstruction.
• All-angle AI video calling.
• 90-degree and 180-degree AI SpinShot modes.
• Motion synchronized to music.
• Cinematic imaging technology developed with ARRI.

These are official product claims and demonstrations. Camera quality, motor durability, stabilization performance, tracking accuracy, acoustic noise, and power consumption will need independent testing on final commercial hardware.

Image credit: HONOR and GSMA/MWC.

Cinematic Imaging Through the ARRI Collaboration

HONOR has also entered a strategic technical collaboration with professional cinema-camera company ARRI.

The companies say ARRI image science, color expertise, visual style, and production knowledge will be incorporated into mobile imaging, with early results appearing on Robot Phone.

This is separate from the agent-workflow story, but it supports a practical use case for the mechanical camera: content creation.

A motorized camera that can track a subject and stabilize movement may allow a creator to appear in a shot without holding the phone or relying on another operator.

The value of the collaboration will depend on how ARRI-related color and imaging features appear in the final device. A partnership announcement alone does not establish that the phone will match professional cinema hardware.

Image credit: HONOR, ARRI, and GSMA/MWC.

HONOR’s “One Primary Agent, Many Specialists” Architecture

HONOR describes the future intelligent-device ecosystem as a structure with one primary personal agent coordinating many specialist agents and devices.

The primary agent would stay close to the user, maintain authorized context, and understand preferences. It would then call specialist capabilities when needed.

A travel request might involve:

• A personal agent that understands preferences.
• A transport service.
• A hotel-search agent.
• A calendar service.
• A payment service.
• A navigation system.
• A cloud reasoning model.
• Local device sensors.

This architecture attempts to connect three categories that are usually fragmented:

1. Devices
2. Services
3. Data

HONOR’s earlier Agentic OS materials described coordination across personal devices, edge services, and cloud intelligence. The WAIC announcement emphasizes a portable primary agent connected to specialized terminals, vertical ecosystems, and cloud models.

The concept is attractive, but it introduces difficult governance questions.

Who controls the user profile? Which agent can access which data? How are competing services selected? Can one provider see the context passed to another? How does the user revoke access? What happens when specialist agents disagree?

These issues will shape whether a multi-agent ecosystem feels helpful or intrusive.

From Tool-Like AI to a “Partner” Experience

HONOR frames its AI strategy around Augmented Human Intelligence, or AHI.

The company’s message is that AI should expand human ability rather than replace people. In this framework, intelligence is described through both IQ—problem-solving capability—and EQ—the ability to understand and respond with sensitivity.

Robot Phone is presented as a physical expression of that idea.

A conventional assistant responds with words, notifications, and screen changes. Robot Phone can orient toward the user, maintain attention, move with music, and use its camera as a form of body language.

HONOR CEO James Li has described the desired transition as one from tools to partners and from commands to intention-driven interaction.

Terms such as “self-awareness,” “life-like intelligence,” and “human-like life form” appeared in the conference discussion. These phrases are philosophical and promotional, not evidence that the product possesses consciousness or biological qualities.

It is more accurate to describe the device as an engineered system designed to produce partner-like interaction through memory, multimodal sensing, planning, software execution, and physical movement.

What Must Be Solved Before Robot Phones Become Mainstream

Robot Phone offers a vivid picture of embodied mobile AI, but the commercial category still faces major questions.

Mechanical Reliability

A motorized camera introduces moving parts that can wear, collect dust, suffer impact damage, or become misaligned.

Battery Life

Continuous sensing, model inference, camera tracking, and motor movement may consume significant power.

Thickness and Weight

The camera mechanism requires internal space. Buyers will need to decide whether its benefits justify a potentially larger or heavier device.

Privacy

A camera that can turn independently needs visible indicators, clear permission controls, and predictable behavior.

Safety

The mechanism should retract safely, detect obstructions, stop when movement is blocked, and avoid pinching or striking nearby objects.

App and Service Compatibility

Cross-app automation depends on stable integrations. Interfaces, login systems, and regional service availability change frequently.

Error Recovery

A useful agent must recognize failure, explain what went wrong, preserve task state, and let the user take over.

Payment and High-Impact Actions

Purchases, bookings, messages, registrations, and account changes should require appropriate confirmation.

Transparency

Users need to know which model, service, account, memory, and permission is being used for each action.

Price

The final price will determine whether Robot Phone is a specialist creator device, a premium flagship, or the beginning of a broader product category.

Announced Features Versus Information Still Missing

Confirmed or Officially Demonstrated	Still Awaiting Full Commercial Detail
Robot Phone exists as working hardware	Final retail price
Reservations opened in China	Complete global release schedule
Two new Chinese-market colors announced	Official English color names
Compact 4DoF gimbal system	Full dimensions and weight
Three-axis mechanical stabilization	Battery capacity and endurance
AI object tracking and SpinShot demonstrations	Final camera specification set
Agentic OS framework	Exact feature availability at launch
Qwen-based collaboration with Alibaba	Public benchmark methodology
OpenClaw platform concept	Complete developer API and policy
ARRI technical collaboration	Scope of ARRI features in shipping software

HONOR’s official global product page explicitly says that shown features remain under preparation and that the final commercial experience may differ.

That note is important. Robot Phone is closer to market than a static concept, but some parts of the ecosystem are still being completed.

FAQ

What is the HONOR Robot Phone?

HONOR Robot Phone is a smartphone with a deployable, motorized camera gimbal and embodied AI features. HONOR positions it as a new device category combining a phone, AI agent, robotic motion, and cinematic imaging.

Is HONOR Robot Phone available to buy?

HONOR announced that reservations had opened through its channels in China on July 18, 2026. Complete pricing, delivery timing, global availability, and final configurations were not included in the source report.

What is a 4DoF gimbal?

A four-degrees-of-freedom system can move across four independent motion dimensions. HONOR uses the mechanism for camera stabilization, tracking, expressive movement, and embodied interaction.

What can Robot Phone’s AI agent do?

HONOR demonstrated cross-app tasks such as ordering a cake, arranging transportation, and booking a karaoke room. The system is also intended to understand multimodal input, use memory, call tools, recover from errors, and coordinate digital and physical actions.

What is Agentic OS?

Agentic OS is HONOR’s proposed operating-system architecture organized around user intent and tasks rather than isolated applications. It combines multimodal interaction, proactive planning, execution, and cross-device agent coordination.

Does Robot Phone use Alibaba’s Qwen model?

HONOR and Alibaba have announced a jointly optimized terminal-model solution based on Qwen for mobile scenarios, and the source report says it has been implemented on Robot Phone. The exact model versions, local-versus-cloud split, and final deployment details have not been fully disclosed.

What is OpenClaw on Robot Phone?

HONOR describes OpenClaw as an open platform that can turn the phone’s gimbal into an agent-controlled actuator. Public documentation for its complete APIs, permissions, and release status was not identified at publication time.

Are all demonstrated Robot Phone features final?

No. HONOR’s official product page states that the displayed features are still under preparation and that the final interface and functionality are subject to the commercial release.

Related Tools

• HONOR Robot Phone: HONOR’s official product overview for the gimbal, embodied interaction, tracking, imaging, and ARRI collaboration.
• HONOR Agentic OS: HONOR’s official explanation of its intent-centric mobile operating-system architecture.
• HONOR Developer Services: The developer portal for HONOR device services, applications, documentation, and ecosystem integrations.
• Qwen: Alibaba’s multimodal model family used across cloud, mobile, and physical-AI applications.
• Alibaba Cloud Model Studio: A platform for developing and deploying applications with Qwen and other generative AI models.
• ARRI: The professional cinema technology company collaborating with HONOR on Robot Phone imaging.

Related Links

• HONOR’s Official WAIC 2026 Announcement: Official details on reservations, colors, Agentic OS, Qwen collaboration, demonstrations, and OpenClaw.
• HONOR Robot Phone Official Page: Product demonstrations and official descriptions of the 4DoF gimbal, tracking, movement, and imaging functions.
• HONOR Agentic OS Technical Direction: HONOR’s four-part definition of its next-generation mobile operating system.
• HONOR MWC 2026 Robot Phone Announcement: The earlier global introduction of Robot Phone and its embodied-AI design.
• Alibaba Cloud on Qwen’s AI Hardware Ecosystem: Alibaba’s account of Qwen adoption in hardware and its work with HONOR.
• HONOR and ARRI Collaboration: Official information on the companies’ mobile-imaging partnership.
• WAIC Official Website: Official information about the World Artificial Intelligence Conference.

Summary

HONOR used WAIC 2026 to move Robot Phone closer to commercial reality, opening reservations in China and demonstrating how a motorized camera can serve as both an imaging system and a physical interface for an AI agent.

The device is built around a compact 4DoF gimbal, three-axis stabilization, AI tracking, expressive motion, Agentic OS, OpenClaw, and a Qwen-based mobile solution developed with Alibaba. HONOR also wants the phone to coordinate long-horizon tasks across applications rather than simply answer questions.

The most difficult work will be reliability, not spectacle. Cross-app execution, error recovery, permissions, mechanical durability, privacy, battery life, and transparent confirmation will determine whether embodied AI becomes a useful phone capability or remains a premium novelty.

Robot Phone’s real significance is not that its camera can move—it is that HONOR is trying to make physical movement, software execution, and multimodal reasoning part of one mobile-agent system.