By Da Cheung
The embodied AI industry — a sector focused on giving artificial intelligence physical robotic bodies — is booming, chaotic, and heavily funded. After a century of science fiction imagination, the dream of intelligent robots living among humans feels like it is nearing an explosive tipping point of realization. Yet, as startups burn cash in the scramble to build the perfect robotic worker, they are placing their bets on entirely different technological tracks.
On June 15, Shanghai-headquartered ACE Robotics announced its Angel+ round of financing, pushing its cumulative funding for the year to hundreds of millions of US dollars. But even as investors throw cash at the industry, it is facing a profound underlying challenge: the hardware, data, and models required to build a functioning robot remain deeply fragmented, making it difficult to judge who is actually on the right path.
The simulation shortcut
One of the largest hurdles in robotics is the staggering lack of training data. AI text models can scrape the internet, but robots need physical interaction data. According to Zhang Minying an algorithm expert at Alibaba Cloud, training a breakthrough embodied AI requires about 100 billion hours of data, while the global supply of high-quality, real-world physical interaction data currently sits at roughly 500,000 hours.
To bridge this massive gap, ACE Robotics has turned to digital fabrication. The startup partnered with a lab from the Chinese University of Hong Kong to launch Kairos-HomeWorld, a unified framework that generates virtual environments. The project created digital replicas of 300,000 authentic Chinese residential layouts, complete with local architectural traits like enclosed kitchens. Instead of deploying expensive physical machines, the company allows its virtual robots to train inside these digital homes.
Wang Xiaogang, founder of ACE Robotics, told 36Kr that the embodied AI supply chain is too long for any single company to master entirely, so his strategy focuses on securing scalable commercial scenarios first. The company’s approach to household training shares the same goal as competitors like American startup Figure AI — which has gathered data from more than 100,000 real physical homes — but ACE can generate virtual scenes based on existing real-world floor plans to drastically lower costs and bypass physical limitations. Even so, analysts from EmbodiCamp caution against relying purely on this generated data, warning that simulations still struggle to replicate complex physics accurately.
Rethinking how robots move
While some companies focus on where robots train, others are reconsidering how they process instructions. Most of the industry relies on Vision-Language-Action, or VLA, models — AI systems that directly translate visual input and text commands into physical robotic movements. These models typically predict actions frame-by-frame, calculating microscopic adjustments every fraction of a second.
X Square Robot, another emerging startup, argues that this conventional method is inherently flawed. When an AI memorizes a rigid sequence of micro-movements, a slight change in the environment, like a differently shaped table, can cause the robot to fail entirely.
In response, the company released WALL-WM, billed as the world’s first “event-level prediction” world model. A world model is an AI system that builds an internal understanding of the physical environment’s logic to predict future outcomes. Instead of planning a rigid trajectory, WALL-WM processes semantic events. If told to grab a cup, the model imagines the physical state of the world at the exact moment the cup is grasped and then generates the fluid movements required to reach that state. According to the company, this event-centric approach allows robots to adapt more seamlessly across unpredictable environments.
Selling shovels in the AI gold rush
Rather than competing to build the smartest robotic brain, Nasdaq-listed Hesai Technology (2525.HK) (HSAI) wants to monopolize the senses. Already a dominant, profitable player in the automotive sensor market, the company recently announced a strategic pivot from “spatial perception” to “spatial intelligence”.
The company unveiled the “Picasso” chip, a 6D full-color LiDAR, a remote sensing method that uses pulsed lasers to map 3D environments. Traditionally, when used for automobiles, LiDAR only mapped spatial coordinates, leaving cameras to figure out colors and textures. Hesai’s new chip natively fuses distance and color data at the hardware level, allowing a robot to instantly recognize both the shape and the semantic identity of an obstacle without processing delays.
Additionally, the company introduced Kosmo, a portable AI hardware device designed to quickly scan the real world into high-quality, editable 3D training environments for robots. By providing the essential sensory hardware and data collection tools, Hesai positions itself as the “shovel seller” in the AI gold rush. The company avoids the risky bet of predicting which robot model will eventually dominate, opting instead to supply the foundational infrastructure every robot will need.
It’s almost impossible to predict which technological path will ultimately win out. As companies push distinct approaches in simulation, event-level reasoning, and sensory hardware, the industry is poised for an extended period of cash burn. It may require several rounds of brutal industry consolidation before the dust settles and a clear winner emerges.
Sources:
