Gasgoo Munich-The World Artificial Intelligence Conference (WAIC) sent an unprecedented signal: artificial intelligence is moving beyond merely "seeing and speaking" in the digital realm to "executing and controlling" in the physical world. Capital, as always, smells the opportunity first. Global funding for physical AI topped $6.4 billion in just the first quarter of 2026, with more than $100 billion pouring into the sector over the past 18 months. The market is placing a heavy bet on this paradigm shift from "virtual" to "real."
Physical AI is not simply "robot plus AI." As Zhang Lihua, a professor at Fudan University, points out, perceiving space is not the same as completing tasks. A vast chasm remains—spanning state modeling, physical deduction, real-time control, and safety verification. Mistakes in the digital world can be rolled back, but an agent's misjudgment in a physical setting results in irreversible, real-world damage.
It is in this field that Wang Xiaogang, chairman of ACE Robotics, has proposed the first principles of the Embodied World Model. The model's core value lies in extracting high-density, critical information from redundant multimodal inputs to anchor "control-sufficient states," thereby minimizing the cost of action for robotic agents in uncertain real environments. The underlying logic is straightforward: the richer the key information a robot masters, the more precise its judgment of control-sufficient states, and the lower the cost of action when executing tasks.
Image Source: ACE Robotics (same below)
The Brain: How to Sharpen Insight?
For a long time, the "brain" of embodied intelligence has suffered a natural split: representational models excel at abstract reasoning but lack a sense of reality; generative models can describe pixels but don't understand physical laws; and interactive models are immersed in 3D simulation but struggle with closed-loop control. This "imbalance" in capabilities has prevented robots from forming a complete intelligent loop of "understanding, deduction, execution, and reflection" in the real world.
Daxiao's released Kairos 3.1 world model stands out for its native unified architecture. It is not a patchwork of three independent modules but relies on a hybrid Transformer and a shared mixed-attention mechanism to compress multidimensional embodied data—including visual observation, language instructions, force-tactile states, and policy trajectories—into a single unified latent space. Here, generative intelligence constructs the representation of the physical world, physical intelligence deduces causality and parallel universes, and cognitive intelligence handles long-horizon task decomposition. These three are deeply intertwined, ensuring that every "imagination" by the robot accurately predicts the real-world impact of its actions.
This is not just a smarter "brain," but an "action-oriented" system capable of real-time edge inference. Relying on the self-developed KairosRT engine, its inference latency on the NVIDIA Jetson Thor platform is just 125 milliseconds. Compared to the 648-millisecond latency of the similar Cosmos 3 Nano model, this represents a 52-fold increase in efficiency. In real-world home tests, robots equipped with this model can autonomously break down more than a dozen laundry steps. When facing failure, they can pinpoint the node and "restart sub-tasks," even iterating from a three-finger grip to a four-finger strategy when failing to open a fridge door. This "reflective self-evolution" capability has helped Kairos 3.1 secure 12 SOTA (State-of-the-Art) titles in global authoritative evaluations for spatial understanding and navigation decision-making.
Data: Value Lies in 'Quality'
AI's "enlightenment" depends on data, yet the logic of data for physical AI differs fundamentally from that of language models. Daxiao introduced the "Law of Information Density" at the conference: the value of data lies not in sheer volume, but in its ability to alter action outcomes. Low-dimensional L1-L2 data can only support basic perception; only by entering L5-level data—which incorporates 3D force-tactile sensation, failure recovery trajectories, and open-scenario variables—can world models truly move beyond "description" to achieve generalization and self-evolution.
The human-centric environmental data collection solution 2.0, released based on this law, redefines the underlying paradigm of data production. Its core is a closed-loop system that "collects fully, labels finely, and uses smartly." The ACE Ego Kit achieves high-fidelity, full-dimensional data collection through a "head-hand-chest" integrated lightweight suite. Notably, the ACE Sense Glove is an industry-first tool integrating 3D force-tactile and motion capture information, boasting 0.01N force sensitivity and less than 2° joint angle error. It synchronizes over 20 heterogeneous sensors with an error of less than 1 millisecond, precisely recording interaction data embedded with physical causality.
In the labeling phase, the generative 4D hand motion capture paradigm ACE-ViDiHand overcomes challenges of occlusion and rapid movement. With a frame-level accuracy of 0.997, it increases action smoothness by 4.8 times. Ultimately, the open-source embodied data foundation, ACE Ego Matrix, achieves the "four unifications" of spatial coordinates, body structure, and action timing, breaking hardware barriers so heterogeneous robots "speak the same language." Daxiao has further open-sourced the L5-level home complex task dataset ACE-Data-0 to the industry. This is akin to providing a high-precision "map" to a sector still in exploration, accelerating the entire physical AI industry from "workshop-style" training to industrial production.
Implementation is the Hard Truth
The ultimate testing ground for technology is the real world. If physical AI cannot solve commercial pain points, it remains nothing more than an expensive laboratory toy. The three standardized industry solutions released by Daxiao demonstrate precisely the sharpness of its conversion from technological potential to industrial kinetic energy.
In the instant retail scenario, the "Xiaoman" solution directly addresses challenges of multi-SKU management, high-density narrow warehouses, and round-the-clock order peaks. The fulfillment robot W1 can navigate aisles as narrow as 75 centimeters. With a load-to-weight ratio of less than 2 and extreme force control of ≤1N, it effortlessly handles items ranging from flexible packaging to rigid goods. By bridging the full chain of "mechanical control—world model—order warehousing," "Xiaoman" has achieved daily lightweight deployment and self-evolution in real-world scenarios like Shaomai Gou and PetroChina convenience stores. The plan is to expand to 10,000 retail stores within the next two years. This is no longer mere showmanship; it is a direct reshaping of the retail cost structure.
In the hotel service scenario, the "Xiaoxin" solution targets the persistent pain point of nighttime laundry. Facing demand that accounts for 80% of the daily total, the robot autonomously links the entire process of "collecting, loading, washing, and folding," transforming non-standard laundry rooms into unmanned, closed-loop operations and drastically slashing night-shift labor costs. In open scenarios, the "Xiaotu" quadruped solution demonstrates "one-brain-many-forms" versatility. By adapting a general intelligent brain to multi-form quadruped robots, it has achieved 7×24-hour uninterrupted autonomous operation in settings such as Shanghai Riverside urban governance and park security.
Even more noteworthy is the debut of "PHYSICAL IQ," the industry's first unified evaluation benchmark for embodied physical intelligence, and the "World Model Cloud Ecosystem" jointly built by Daxiao with computing giants like Baidu Intelligent Cloud and Alibaba Cloud. These developments signal that the competition in physical AI is shifting from a battle over individual technical specifications to a struggle for ecosystem synergy and the right to define standards.
The "enlightenment" of physical AI is not the sudden flash of a single algorithm, but the "trinity" emergence of world models, high-density data, and vertical scenarios within a commercial loop. From the native unified architecture of Kairos 3.1 to the new data laws defined by the environmental collection solution 2.0, and down to the solid implementation of the "Xiao" series solutions, Daxiao has outlined a clear evolutionary map for physical AI. It tells us that intelligence can only move from the digital "illusion" to the industrial "reality"—and usher in that "enlightenment moment" for the entire era—by anchoring to the causality of the physical world, bearing the cost of real action, and continuously evolving through self-reflection.
