In physical AI, the real bottleneck is not how well a machine moves but whether it can continuously perceive and understand the physical world around it — and the industry has barely started to build the data pipeline that would make that possible at scale.
The robotics industry has spent a decade arguing about which robot moves best. That debate is becoming obsolete. In physical AI — the field of teaching machines to perceive and act in the real world rather than generate text and images — the actual competitive fault line has shifted. The companies that will win the next phase are not the ones with the most elegant gait or the largest model; they are the ones that crack the continuous supply of real-world perception data that lets a robot actually understand the messy environment it has been dropped into.
That matters because the public conversation around humanoid robots, autonomous vehicles, and warehouse automation is still anchored to movement. Industry coverage tends to benchmark gait, dexterity, and benchmark scores, treating the perception problem as essentially solved. The deployment record tells a different story. As robots leave staged demos and enter open environments, perception failures rather than locomotion failures explain most of the gap between controlled and real-world performance.
The mechanism is straightforward once stated. A perception model is only as good as the data that trains it. In physical AI, training data has to be captured from the actual physical world rather than scraped from the internet: cluttered kitchens, dim warehouses, uneven sidewalks, shifting crowds. That data is expensive to collect, slow to label, and constantly drifting as the world itself changes. No single lab can scale that pipeline alone.
One company that has positioned itself squarely in this layer is LeDong Robotics, which listed on the Hong Kong Stock Exchange main board in late April 2026. Founded about a decade ago as a maker of vacuum robots, LeDong's Hong Kong listing prospectus describes a business that has progressively migrated its visual and spatial sensing work outward, into logistics robots, hospitality service robots, and emerging humanoid and quadruped platforms. A QbitAI company feature following the listing frames the entire company as a bet that perception, not movement, is the binding constraint in physical AI.
That framing is best read as a company narrative before it is read as an industry verdict. The QbitAI piece is a sponsored-style profile anchored on chairman quotes, and it floats a trillion-dollar market claim for the company's addressable opportunity that no independent sizing supports. Other Chinese financial outlets, including iFeng, have corroborated the listing event itself, but the market sizing remains the company's own projection. Treat the thesis as plausible, the numbers as unverified.
Independent analysis leans toward the same structural conclusion even when it is uninterested in LeDong specifically. A long-form Zhihu analysis of the embodied AI stack reaches a similar diagnosis from a different direction: as model size stops being the binding constraint, attention shifts to data infrastructure and the operational layer where perception models meet the world. LeDong is one of several players positioned in that layer. Visual-LiDAR fusion vendors, robot-data-platform companies, and the embodied-AI model teams themselves are all making competing bets on who owns it.
The competitive picture is also more defensive than the LeDong profile suggests. A Digitimes piece on Taiwanese motion-component maker TBI Motion published the same week treats the humanoid push at adjacent suppliers as a survival hedge against commoditized motion components, not a moonshot into a trillion-dollar future. That counterpoint matters. If upstream component makers are hedging rather than committing, the perceived gold rush in humanoid robotics is not yet showing up in the supply chain's investment behavior.
The remaining questions are concrete and answerable. Which perception-stack vendor can demonstrate a real-world data flywheel, where each deployed robot generates the labeled data that trains the next generation of models, rather than a one-off demo pipeline. How the industry handles long-tail scenarios: the unfamiliar object, the unusual lighting, the crowd pattern the model has never seen. And who owns the resulting data assets, the operator that captured them, the model vendor that labels them, or the customer whose environment produced them.
Those are not questions any single company can answer alone. They are the questions the next phase of physical AI will be judged on, and the companies that win will be the ones that built the perception pipeline before anyone realized motion had stopped being the bottleneck.