On 22 April 2026, the FAIR plus Robotics Full Industry Chain Conference officially opened at the Shenzhen Convention and Exhibition Centre (Futian). Positioned as a world-class conference on robotics development and manufacturing technology, this event focuses on technologies and development resources across the entire robotics industry chain. Through academic exchanges, technical standards workshops and supply-demand matchmaking sessions, it aims to establish an in-person platform for collaboration between technical professionals in the fields of artificial intelligence and robotics, as well as end-users, thereby driving the intelligent upgrading of the robotics industry and the development of its overall capabilities.

Noin showcased its product prototype in person at FAIR plus, demonstrating its ability to perform complex operations under any lighting conditions and against any background; the demo area was bustling with visitors:

It takes about a minute to fold a garment; even if someone tucks the corners away, Xiao Nuo can still manage it:

(All the videos below are actual footage at 1x speed)

Just how difficult is it? Let’s take a look at the reflections on the desk, changes in the person’s shadow, changes during video playback, and changes in the flash—even under the ‘Hell’ difficulty interference mode, Xiao Nuo can still complete the task:

Some members of the audience even pulled clothes out of their bags on the spot to set challenges:

Faced with completely unexpected types of clothing, Xiao Nuo passed the challenge with flying colours thanks to its powerful generalisation capabilities!

Why is Xiao Nuo able to operate reliably even under “hell mode” conditions—with unpredictable lighting, complex backgrounds and unknown objects?

The answer lies in Noin’s unique technical architecture.

Noin hasIn-house developed embodied intelligence large language model和The KnowinDream Data Generation Engine!This engine can generate vast amounts of data in batches featuring “any lighting + any background + any object + any task”, enabling Noin’s in-house large language model to undergo thorough training and anticipate a wide range of extreme real-world scenarios in advance, thereby endowing Xiao No with exceptional environmental adaptability and task generalisation capabilities.

As well as its clothes-folding skills, the Noin robot’s ability to organise items on a desk is equally impressive! Equipped with a proprietary embodied large language model, Xiao Nuo can accurately understand natural language commands and, combined with visual recognition, carry out complex organisation tasks.

At the exhibition, the “You Set the Task, I’ll Do It” interactive session was full of highlights. As members of the audience pulled out their earphones and hair clips, Xiao Nuo performed zero-shot adaptation and organised them with precision, achieving a task completion rate of approximately 98%.

Upon receiving the instruction to “hand me the car keys”, Xiao Nuo thought for a moment and then delivered them with precision:

Faced with the spatial requirement to “wipe only the desk surface between the two pens”, spatial orientation is easily achieved:

Someone else posed this question: “Pass me the purple ear stud in front of you, but you’re not allowed to hold it between your fingers.” My brain went completely blank—what on earth is this question?

After giving it some thought, Xiao Nuo successfully passed the earring on to the audience member who had posed the question—well done!

Why is it possible to achieve “perfect execution of any instruction”?

It is primarily underpinned by two core technologies:

Firstly, KnowinBrain, the foundational platform for embodied cognitive large models — by utilising a 1–5 Hz execution feedback mechanism, the large model is capable of autonomous thinking, dynamically assessing its physical state, and autonomously reflecting upon and correcting itself, thereby establishing a closed-loop system linking perception, planning and execution.

Secondly, the generative action encoder-decoder, Action Tokenizer, is pre-trained on video data generated by KnowinDream. It possesses the ability to generalise action representations, ensuring smooth operation in unknown tasks and environments; furthermore, it achieves full alignment with both semantic and visual information, truly enabling it to “do what it sees and hears”.

Noin has abandoned the technical approach of mimicking large language models; when carrying out tasks, it first seeks to understand the real world before generating the desired actions, rather than simply imitating them!

The exhibition was bustling with visitors, and the market’s anticipation for smart home technology was clearly evident.

Noin, developed in-houseKnowinDream Synthetic Data EngineOvercoming the bottleneck of data generalisation in domestic settings, by relying onKnowinBrain, a fully in-house developed full-stack solution, and interactive feedback technology, to accelerate the practical implementation of Level 3 autonomous household robots.

Moving from the laboratory to in-person exhibitions to engage with the public marks a key milestone in the successful validation of our generalisation capabilities; in the future, we will continue to focus on all aspects of home life, committed to bringing intelligent robots into every household.

KNOWIN Embodied Foundational Model Architecture: A Full-Stack Generative and Generalisable Embodied Technology Framework

<