---
title: "The intelligence shortcomings of robots are being filled in by the Da Xiao World Model."
type: "Topics"
locale: "en"
url: "https://longbridge.com/en/topics/43905850.md"
description: "The narrative of embodied AI mass production has been a recurring theme this year, captivating both capital and the market. This sentiment was further amplified by Unitree's listing, which rapidly expanded its imaginative potential. As 整机 prices continue to drop, expectations along the supply chain have heated up; from actuators to lead screws and dexterous hands, every component offers new avenues for valuation imagination. However, Unitree's market cap halving in less than a month has prompted the market to re-evaluate whether core hardware, sales volume, and mass production expectations alone can sustain such valuations. Starting in Q3, industry focus has shifted noticeably toward the "robot brain," with world models entering a period of intensive releases. On September 9..."
datetime: "2026-09-15T10:09:04.000Z"
locales:
  - [en](https://longbridge.com/en/topics/43905850.md)
  - [zh-CN](https://longbridge.com/zh-CN/topics/43905850.md)
  - [zh-HK](https://longbridge.com/zh-HK/topics/43905850.md)
author: "[朝阳资本论](https://longbridge.com/en/profiles/26763750.md)"
generator: "portal-rs"
---

# The intelligence shortcomings of robots are being filled in by the Da Xiao World Model.

The narrative of mass production for embodied AI has been a constant theme this year, one that capital loves to hear and the market eagerly chases, especially with Unitree's listing rapidly expanding its imaginative potential.

**As 整机 prices continue to drop and supply chain expectations heat up, new valuation imaginaries can be found in everything from actuators to lead screws and dexterous hands for a single robot.**

However, the fact that Unitree's market cap halved in less than a month has forced the market to re-examine how much valuation can truly be supported by the 本体 alone, sales volume, and mass production expectations.

Starting in Q3, industry attention on the "robot brain" has noticeably intensified, with world models entering a period of dense releases.

On September 9, JD.com released the JoyAI world model; on the 10th, Unitree open-sourced WLA, adding world modeling capabilities, while AutoNavi released the 3D-native urban world model ABot-Earth 0.7; on the 11th, Ant Lingbo continued to open-source LingBot-World 2.0.

A company established for just over a year has also quickly broken out in this race.

Recently, Daxiao Robotics, in collaboration with Nanyang Technological University's S-Lab, released Puffin-World, attracting significant attention among the densely released world models.

**Puffin-World directly defines physics, geometry, and appearance as three native 3D world states, not only generating visuals but also understanding gravity, spatial structure, and predicting future world states.**

In embodied AI, intelligence has long been the key differentiator.

**Embodied AI Exposes Intelligence Shortcomings**

Earlier this year, Turing Award winner Yann LeCun, a key driver of the world model route, stated that many of the "kung fu" stunt moves performed by humanoid robots are pre-calculated. **The industry's biggest secret is that no company truly knows how to make robots smart enough to be useful; existing robots still haven't even reached the common sense level of a house cat.**

Data from IDC shows that global shipments of humanoid robots exceeded 18,000 units in 2025, with Chinese manufacturers accounting for nearly 95%. **Yet, nearly 90% of these products still flow into scenarios like entertainment performances, educational research, data collection, and guided exhibitions, with only 9.2% truly entering industrial manufacturing.**

By the first half of 2026, global shipments further exceeded 22,000 units, yet smart manufacturing and warehousing logistics combined accounted for only about 18%.

Taking Unitree as an example, in the first three quarters of 2025, 73.6% of its humanoid robot revenue came from scientific research and education, with industry applications at only about 9%.

A large number of robots haven't entered factories to work; instead, they remain in labs, classrooms, and exhibition stages performing tricks.

**To bridge the gap between theory and reality, world models are one of the most important technical routes.**

A world model essentially allows AI to build an internal simulation of the real world. Robots need to know not only what they see but also understand spatial relationships between objects, basic physical laws, and how the environment might change after an action is taken.

If robots can simulate consequences internally, cross-scenario generalization truly has a foundation.

Currently, capital's attention appears to be shifting.

Data from IT Juzi shows that in the first eight months of 2026, the domestic embodied AI sector completed 466 financing rounds, reaching $124.51 billion. **Among them, financing rounds for embodied brain systems accounted for 38.8%, surpassing the 21.1% for humanoid robot 整机， becoming the sub-sector with the most financing rounds.**

By August, of the 56 financing rounds tracked by Gaishi Embodied Intelligence, nearly half flowed into companies related to brain systems.

Thus, world models have moved from a frontier technical route to the center of the table.

Daxiao Robotics' launch of Puffin-World at this moment hits precisely this window.

**Daxiao Bets on World Models**

Daxiao Robotics is a very young company in this wave of embodied AI entrepreneurship.

The company was established in July 2025, with Sun Wen Technology co-founder Wang Xiaogang serving as Chairman and Tao Dacheng as Chief Scientist.

In the first half of 2026 alone, Daxiao raised hundreds of millions of dollars, with participation from Ant Group, Geely Capital, Dascent Capital, Shenzhen Capital Group, and others.

**In just over a year since its establishment, Daxiao has continuously launched multiple generations of robot foundational models and world models. The recent sensation in the industry is the joint release of Puffin-World with NTU S-Lab.**

Unlike many robot companies that start from the 整机 end, Daxiao placed its focus on the model layer from the beginning.

In June this year, Daxiao released the Kairos world model tech stack, upgrading to Kairos 3.1 in July. Kairos addresses problems closer to the robot's action end, putting visual, language, force-touch information, and motion trajectories into a single model to deduce subsequent changes based on the current state before generating actions.

**This route anchors the longest-range tasks that humanoid robots find hardest to solve.**

It's not difficult for a robot to grab a cup once; the difficulty lies in continuously completing a sequence of actions like opening a door, picking up an object, moving, and placing it. If any step deviates, subsequent actions must adjust accordingly. If the model can only execute pre-learned action sequences, the longer the task, the more errors accumulate.

Daxiao has previously focused on such issues. Kairos 3.1 further added long-term memory and edge-side deployment, hoping to allow robots to continuously judge environmental changes during tasks rather than relying on fixed scripts for every step.

**The problem is that to achieve task execution, there must first be a sufficiently accurate world in the mind.**

**Puffin-World is the answer.**

Many past world models focused mainly on RGB video prediction, generating the next frame based on previous visuals. This route excels at generation but ignores much information crucial to robots, such as gravity direction, camera pose, object distance, and spatial structure.

Puffin-World splits physics, geometry, and appearance into three native states for unified modeling.

The physics state handles gravity direction and camera pose, the geometry state describes depth and spatial structure, and the appearance state corresponds to the final visual image seen.

**Based on this representation, Puffin-World uses one model to simultaneously complete four types of tasks: physical world understanding, free-viewpoint simulation, 3D generation and reconstruction, and closed-loop exploration.**

In particular, closed-loop exploration means the model is no longer just passively generating visuals but can judge where to observe next based on the current environment, continuing to update its understanding of the world. For robots, this is closer to real needs than simply making video generation more realistic.

Data scale has also taken a step forward.

**The Puffin-16M built by Daxiao and NTU S-Lab contains 15 million vision-language-camera triplets and 1 million complex camera trajectories; meanwhile, the team supplemented absolute camera pose annotations for 28 public datasets, covering approximately 44.5 million images.**

From public evaluations, this model has achieved results.

Puffin-World achieved leading results in four categories: camera-to-world understanding, camera-controllable generation, complex trajectories, and 3D world modeling. Among them, all 12 median error metrics for camera understanding were optimal, and 33 out of 36 AUC metrics reached the best or tied for best; in RealEstate10K 3D world modeling, it also achieved a PSNR of 17.22 and an LPIPS of 0.318.

Putting Kairos and Puffin-World together, Daxiao's technical route becomes easy to understand.

Puffin is responsible for representing the real world more accurately, while Kairos pushes this understanding further toward task planning and action execution. Combined, they point to a robot brain that can be licensed to the entire industry.

If this path succeeds, the valuation anchor for the humanoid robot industry will shift completely.

**Robot Brains Deserve Independent Valuation**

In the past, capital markets priced robot companies primarily around the 整机。

Whether a company could manufacture robots and deliver stably ultimately determined its revenue scale and valuation space.

**Even if humanoid robots wear the cloak of cutting-edge technology, the underlying valuation logic remains inseparable from manufacturing.**

But when competition shifts to the robot brain, this valuation method is insufficient.

The biggest difference between models and software lies in the fact that one R&D effort can be replicated repeatedly. **If the same robot brain can adapt to bodies from different manufacturers and enter various scenarios like factories, warehouses, and homes, software and models have the opportunity to be separated from hardware, forming a new value layer.**

In fact, capital has already begun attempting to price this possibility.

In January this year, robotic foundational model company Skild AI completed a $1.4 billion financing round, with a valuation exceeding $14 billion.

Skild's goal is to create a universal brain capable of controlling different robots and completing different tasks.

Another company betting on universal robot models, Physical Intelligence, also reached a valuation of approximately $11 billion in its latest financing round.

Behind the valuations of these companies, the market is actually purchasing a platform expectation. **In September this year, Skild disclosed that in just 10 months of commercial deployment, its annualized revenue had broken through $100 million, with over 60 paying customers, applications covering warehousing, factories, data centers, delivery, security, and other scenarios.**

In the future, there may be many robot brands, but the brain responsible for understanding environments, planning tasks, and generating actions does not necessarily need independent R&D.

Similar value migrations have occurred many times in other tech industries.

**After cars entered the intelligent era, while 整车 manufacturing remained important, autonomous driving, operating systems, and software capabilities already possessed independent value.**

**In the semiconductor industry, ARM sells architecture and IP licensing, not chips; Android doesn't charge for phones but for ecosystem position; cloud computing charges by compute power, regardless of who manufactures the servers.**

The difference is that this valuation logic holds even more imagination in the humanoid robot track.

If robots can only perform a few fixed actions, they are 本质上 still automated equipment wearing a humanoid shell; only those capable of understanding environments, planning tasks, and continuous learning approach the universal robots the market truly expects.

World models are entering the critical link of humanoid robot task planning and autonomous decision-making, representing one of the parts with the highest technical barriers and easiest premium formation in the robot brain.

This is also why Daxiao Robotics, in just over a year since its establishment, has quickly gained capital attention.

What Daxiao possesses today is not just a world model.

From foundational models and data bases to continuously expanding body adaptation and scenario connections, these capabilities are being integrated into a single system. Looking at any single item alone makes it hard to explain the full value of a robot brain.

**What truly deserves capital to price early is how large a combination space these capabilities can form in the future.**

From this perspective, Daxiao is no longer competing for leadership in a single model but for the layer of capability most likely to be independently priced in the next-generation robot industry.

As humanoid robots move from hardware competitions to intelligence competitions, world models, data, and cross-body adaptation capabilities are likely to transform from 附属 items of 整机 into core assets determining product boundaries and industry value.

At that time, it will truly enter the home court of Daxiao and others.

Source: Chaoyang Capital Theory

---
> **Disclaimer: This article is for reference only and does not constitute any investment advice.**