Home
Blog
When a 2021 Robot Gets a 2026 AI Brain Ubitus × Kebbi: A Physical AI Experiment

When a 2021 Robot Gets a 2026 AI Brain Ubitus × Kebbi: A Physical AI Experiment

26-09-03

Author: Bull, BD & MKT Director | Editor: Helen, Deputy Marketing Manager

Can a service robot that has already been in operation for years find new possibilities in the era of generative AI?

When people talk about AI robots, attention often goes to newer chips, greater computing power, or newly designed humanoid robots. Yet exhibition venues, campuses, shopping malls, and service environments are already filled with robots equipped with screens, microphones, speakers, cameras, and movable joints. Limited by the hardware and software designs of their time, however, these robots typically execute fixed commands, play prerecorded content, or interact through predefined workflows.

So we wanted to conduct a simple experiment:

What happens if we keep the same body, but give it a new AI brain?

For this experiment, we used the Kebbi Air H201 as the foundation, retaining its existing sensing, display, audio, and motion capabilities. We did not replace the motherboard or attempt to run a large model directly on the robot. Instead, through a lightweight Android app, we connected real-time speech recognition, LLM-powered conversation, text-to-speech, and robot motion control.

The result was more than simply “it can answer questions now.”

The same robot can now detect when someone approaches, greet them proactively, understand natural language, and engage in multi-turn conversations. While responding, it can display subtitles, blink, lip-sync, wave, or dance. If the user starts speaking again or touches the screen, the robot can interrupt its current response and return to listening.

Its body has not changed. But the way it interacts with people has.

From Fixed Commands to Understanding Human Intent

One of the most common limitations of traditional voice robots is that users have to adapt to the machine. They may need to say specific keywords, use fixed sentence patterns, or follow predefined steps. A change in phrasing, a pause in speech, or background noise can easily cause the interaction to fail.

The new AI interaction model reverses this relationship. Users no longer need to memorize commands. They can simply express their needs in natural language. The system continuously processes speech recognition results, determines whether the user has finished speaking or is calling the character, and retains conversational context so follow-up questions can continue naturally. To reduce unintended activation, the system can also be configured to respond only when a designated character name is detected.

More importantly, the interaction is no longer a one-way process in which the robot must finish playing an entire response. If the user speaks again or touches the screen, the system stops the current speech and motion and returns to listening.

This ability to be “interrupted” may seem like a small feature, but in practice, it can determine whether people perceive the robot as a genuine interactive partner.

At this point, the robot is no longer simply executing predefined commands. It is beginning to understand:

What does the person actually want it to do?

From AI Responses to Physical Action

Simply moving a chatbot onto a robot’s screen does not fully utilize the value of having a physical body.

That is why we designed the LLM not only to generate text responses, but also to propose corresponding performance intentions.

For example, when saying “Nice to meet you,” the robot can wave at the same time. When explaining left and right directions, it can raise the corresponding arm. When discussing something cheerful, it can accompany the response with appropriate expressions and movements.

Every action must first pass through a safety whitelist on the app side before being executed through the robot SDK.

This is where an important distinction begins to emerge.

Typical Generative AI outputs are: Text / Image / Audio

When AI enters the physical world, its outputs begin to include: Speech / Expression / Motion / Device Action

In other words, the core question is no longer simply “What content did AI generate?” but:

Can AI translate human intent into executable physical actions?

The process can be simplified as:

Human Intent → Perception → Reasoning → Action Planning → Physical Action → Feedback

This is the minimal Physical AI loop we wanted to explore through the Kebbi experiment.

The Brain and the Body Do Not Have to Age Together

Our architecture uses a “Thin App + AI Gateway + Robot SDK” approach.

Kebbi continues to handle what it already does well: capturing audio, playing sound, displaying expressions, and performing physical movements. More computationally intensive tasks—including STT, LLM, and TTS—are handled by external AI services.

Importantly, the Android app is not directly tied to a single model or cloud service. Instead, it connects to an AI Gateway through standardized interfaces. If the speech recognition model, LLM, or voice model needs to be replaced in the future, the entire robot-side application does not need to be rewritten.

This leads to another concept we believe is important:

The Brain and the Body Can Evolve Separately.

Hardware may remain in use for five, eight, or even more years. AI models, however, may make significant advances every few months.

If the two are tightly coupled, the robot’s intelligence is effectively limited to the moment it leaves the factory. But if sensing and motion remain on the device while intelligence resides in an AI layer that can continuously evolve, the hardware lifecycle and AI lifecycle can be separated.

This means:

LLMs can be replaced
STT / TTS can be upgraded
RAG knowledge can be continuously updated
Personas can be reconfigured
Languages and characters can be adapted for different markets
The same hardware can serve entirely different use cases

The current prototype already supports Traditional Chinese and Japanese demonstration modes. Switching modes changes more than the interface language—it also switches the STT, LLM prompts, TTS, character name, wake-up method, and front-end state.

What is being extended, therefore, is not merely the lifespan of the hardware.

It is the length of time during which the device can be redefined for new purposes.

The Real Challenge Begins When AI Enters the Physical World

Running STT, LLM, and TTS APIs individually is not particularly difficult. The real engineering challenge is making them work together naturally, safely, and reliably over long periods on a physical robot.

Three lessons stood out during development.

1. Perception must be reliable

Text input in a browser is clean. The physical world is not.

Distance, direction, ambient noise, different speakers, and the characteristics of the robot’s microphone can all affect speech recognition. The physical device therefore still requires input thresholds, noise-floor calibration, gain adjustment, and failure recovery.

The first requirement of Physical AI is not whether AI can answer a question.

It is whether it can reliably perceive the real world.

2. AI can propose actions, but execution must remain controlled

LLMs can provide creativity and reasoning, but when controlling physical devices, the model cannot be allowed to call arbitrary SDK functions directly.

Every robot action must first pass through a whitelist and physical validation before being executed by a deterministic control layer.

The principle can be summarized as:

LLM can decide what to do.
Deterministic systems must guarantee how it is safely done.

3. Physical AI must always be able to return to the correct state

Real users interrupt. Networks may briefly disconnect. Microphones can fail. Speech, subtitles, lip-sync, and motion can fall out of sync.

The system therefore requires cancellation, state recovery, watchdogs, and diagnostic mechanisms to prevent the robot from becoming stuck halfway through speaking or moving.

The prototype consequently includes state displays, latency logging, a microphone watchdog, and mechanisms to restore the official interface after the service is stopped.

These may not sound like the most glamorous AI technologies, but they often determine whether a Physical AI prototype can truly move beyond the laboratory.

The Next Step May Not Be One AI per Robot

Kebbi is a good starting point, but we believe an even more important direction to explore is:

One Intelligence Layer, Many Physical Bodies.

Future AI does not necessarily need to belong to one specific robot.

The same AI Agent could understand human intent, preserve context, perform planning, and then use different devices depending on the task.

For example, someone might say:

“Please bring me that item.”

The AI does not necessarily need its own wheels or arms. Instead, it could:

Identify the item → Find an available AMR → Dispatch the task → Wait for the robot’s status → Confirm task completion → Report the result to the user

At this stage, AI is no longer simply a feature of “one robot.”

It begins to function as an Intelligence Layer spanning different robots, sensors, and devices.

Not Every Physical AI Journey Needs to Start with a New Robot

The Kebbi experiment remains an evolving prototype, but it has shown us another possible path for Physical AI.

When enterprises adopt AI in the future, they may not always need to purchase an entirely new intelligent robot first.

Many existing devices already have enough of a “body”: microphones, screens, cameras, speakers, sensors, motors, and network connectivity.

What they may be missing is simply:

An intelligence layer capable of understanding language, maintaining context, reasoning, and translating results into device actions.

This concept is not limited to Kebbi. It could extend to existing service robots, information kiosks, digital signage, AMRs, and other physical devices with controllable interfaces.

In this experiment, we did not build a new robot.

We simply redefined what an existing robot could do.

Kebbi’s screen, microphone, camera, and motors remained unchanged. What changed was how it understood people, how it decided to respond, and how it translated AI reasoning from the digital world into actions in the physical world.

Perhaps this is one direction worth exploring as Physical AI evolves:

The intelligence of the future may not come from a newer body,
but from a brain that can continue to evolve.

Same body. New intelligence.

About Ubitus

As a member of the NVIDIA Connect program, Ubitus leverages NVIDIA’s support and cutting-edge GPU technology to accelerate AI innovation. The company delivers advanced AI solutions, including UbiGPT (a large language model), UbiONE (an AI-powered avatar creation platform), and UbiArt (an image generation tool), providing customized solutions to meet the diverse needs of various industries.

As a cloud gaming pioneer, Ubitus enables Nintendo and other game companies to establish cloud gaming services and supports the global streaming of multimedia content, including interactive and virtual reality experiences.