By mid-2026, embodied intelligence remained one of the hottest sectors in China’s private markets.

But even in a field this hot, Daxiao Robotics, which only debuted at the end of last year, has raised money at a pace that is still unusually fast.

Just now, Daxiao Robotics announced the completion of an angel-plus funding round.

According to people familiar with the matter, the financing round drew participation from multiple investors, including Fortune Capital, Shenzhen Capital Group, Shanghai Sci-Tech Fund, MetaX, Fosun Ruizheng, Huakong Fund and the Lingang New Area Fund. Existing shareholder SenseTime Guoxiang Capital also increased its stake, while Gaojie Capital served as long-term financial adviser.

Including the angel round completed at the end of last year, Daxiao has raised several hundred million dollars in total in just over half a year.

In the increasingly heated embodied AI race, the hundreds of millions of dollars of hot money are betting on a technical path that is rapidly catching fire.

World Model.

More and more people in the industry are coming to a consensus: if robots are to truly enter the real world, simply following commands is far from enough.

It also needs to understand its environment, anticipate changes, and know what may happen next. The world model is the “brain” responsible for exactly that.

Over the past six months, Daxiao has turned in a new answer on the Kairos world model roughly every few months:

Open-source a world model, enable on-device deployment, and launch a digital training ground.

From the model itself to deployment, and then to the training environment, Daxiao’s distinctive world-model approach that integrates understanding, generation, and prediction is gradually taking shape.

Recently, the Kairos world model built on this architecture has topped a series of embodied AI benchmarks, including RoboTwin 2.0, LIBERO-Plus, WorldModelBench Robot and DreamGen.

In a sense, capital is betting on more than just a story about a world model.

Rather, it is a technological path that is being continuously validated and steadily realized.

Capital Bets Big

Before this angel-plus round, Daxiao Robot’s angel financing was led by Ant Group, with participation from Qiming Venture Partners, Kingkey Capital, Hony Capital, Lenovo Capital and Incubator Group, and Hanyuan Asset, the fund of funds affiliated with Shanghai Jiao Tong University. Existing investor SenseTime Guoxiang Capital also kept adding to its stake.

After two rounds of financing, Daxiao brought together more than 15 investment institutions in just half a year.

According to reports, the latest funding round will mainly be used for two purposes: first, to keep advancing iterations of its world model and speed up the deployment of key technologies such as large-scale simulation training and direct control on edge devices.

Second, it will build integrated hardware and software commercial solutions for smart retail, security inspections, cultural tourism, hotels and other scenarios.

Here is a brief primer on what a world model is.

A world model can be understood as a large AI model that predicts “what will happen next.”

Like a large language model predicts the next word, a world model predicts the next frame of a scene and the physical changes within it. For robots, that is effectively a brain that can “imagine” the future.

In Chairman Wang Xiaogang’s view, embodied intelligence is the next era of artificial intelligence, and the “strongest brain” built around a world model is the key to unlocking it.

What has Dayao’s “strongest brain” achieved so far?

Three releases in half a year: from open source to edge devices and then to training.

Looking back over the past six months, Daxiao’s product cadence has barely slowed at all.

In December last year, as the company made its formal debut, it also open-sourced and released the Kairos world model 3.0 (Kairos 3.0).

This is the industry’s first open-source world model that also supports commercial use. It also introduces, for the first time, an integrated “understanding-generation-prediction” architecture, bringing environment understanding, future generation and behavior prediction into a single model framework.

At the same time, Kairos 3.0 has also completed deep adaptation with domestic GPU makers including MetaX, Biren and Sugon, creating a full chain that links “domestic models + domestic computing power.”

Three months later, Daxiao again released the 4B-parameter lightweight Kairos 3.0-4B, achieving the world’s first on-device deployment of a world model.

On NVIDIA’s Thor platform, which is built specifically for robotics, Kairos 3.0-4B has reached a real-time generation speed of 1:1.5. For every second the robot moves forward, it can predict what may happen in the next 1.5 seconds.

In a public demo, a robot powered by Kairos 3.0-4B completed a seven-minute household task in one continuous take: tidying a coffee table, putting away clothes and preparing breakfast, all without human intervention.

Supporting this capability is Daxiao’s continued exploration of model architecture, data systems, and inference efficiency.

At the architectural level, Kairos continues the integrated design approach, combining understanding, generation, and prediction in a single network and directly building a mapping from visual input to action output.

In terms of data, Kairos 3.0-4B was not limited to traditional video training either.

In addition to video data, the model also incorporates reasoning text that describes physical laws, demonstrations of human behavior, and real robot operation trajectories. It learns not just how to do something, but why it is done that way.

On the efficiency side, the team reduced computational complexity from quadratic to linear by developing its own attention operator.

At the same time, Kairos has been adapted to NVIDIA as well as domestic GPU platforms from MetaX, Hygon, and Biren, and supports different robot forms including single-arm, dual-arm, and dexterous-hand setups.

But for embodied intelligence, models and data alone are far from enough.

Robots will ultimately need a sufficiently large training environment before they can enter the real world.

Based on this, a while ago Daxiao again joined forces with CUHK MMLab to release Kairos-Homeworld —

The world’s first unified framework to achieve whole-home 3D generation and object-level full interaction.

Kairos-Homeworld has brought 300,000 real Chinese home floor plans into the digital world, turning them into household environments where robots can train repeatedly.

One notable point is that Kairos-Homeworld is not just a large household-scene dataset; it is also an engine that continuously generates training environments.

With generative simulation, it can keep adding new scenarios at extremely low marginal cost, continuously scaling up the environment.

More importantly, it is the first time Chinese households have been systematically brought into the robot training pipeline.

300,000 real residential floor plans have been turned into training environments, and common features of Chinese homes such as entryways, balconies and enclosed kitchens have been included in large-scale training for the first time.

Through its built-in hierarchical generation framework, the system can automatically handle floor plan generation, furniture layout, scene correction, and physical property assignment.

Parameters such as material, density, and friction coefficient are automatically modeled, allowing robots to train directly on grasping, handling, and manipulation.

From Kairos 3.0 to Kairos 3.0-4B, and then to Kairos-Homeworld, Daxiao has completed a continuous push over the past six months, spanning model development, deployment, and training infrastructure.

And capital keeps betting on this increasingly complete, closed-loop technology stack.

The team that showed up at the buzzer

Daxiao Robotics made its formal debut only in early December last year.

Chairman Wang Xiaogang, a co-founder of SenseTime, has long led the company’s technical efforts in computer vision and autonomous driving.

Chief Scientist Tao Dacheng, a Fellow of the Australian Academy of Science, ranks among the world’s most cited researchers in computer vision and machine learning.

At the same time, the team has also brought together a group of frontier researchers from Nanyang Technological University, the University of Hong Kong, and the Chinese University of Hong Kong, with work spanning ambient intelligence, world models, and embodied foundation models.

On the technical front, Daxiao has built its research around a human-centered embodied intelligence paradigm, putting together an end-to-end system that runs from environment-based data collection to an “Enlightenment” world model and generalized embodied modules.

On the commercialization side, Daxiao has launched its embodied super-brain module A1, which can be adapted to mainstream platforms such as quadruped robots and enables end-to-end visual perception and execution of natural-language commands. Its first deployment targets are the smart retail and hotel sectors.

Following the pace Xiao revealed, the company will also roll out, in quick succession, the latest evolution of the Kaiwu world model 3.0, a scaled-up data collection plan, and an integrated full-stack hardware-and-software solution for retail and hospitality.

— Welcome AI product professionals to build together —

Follow with one click below. Light up the star.