Unitree Robotics founder Wang Xingxing told the 26th China International Industry Fair in Shanghai that the industrial tipping point for embodied intelligence could arrive within one to two years, pegging his forecast to the pace of foundation-model releases5,7. ANALYSIS His remarks landed alongside a new humanoid training model from BeingBeyond, a DeepSeek-led seed round for a robotics startup, and a Shanghai government deployment target, all in the same week, forming the clearest signal yet that China's robotics ecosystem is organizing around a single thesis: general-purpose AI models, not bespoke control stacks, will unlock commercial-scale robots.
Why it matters
The robotics industry has long promised humanoids that work in factories and homes, but generalization, the ability to handle novel tasks without reprogramming, has remained the binding constraint. Wang's claim is that frontier language-and-reasoning models are now closing that gap fast enough to set a calendar. He said his team has already deployed OpenAI's recently released GPT-6 Astra onto robots for task testing, "with strong results". He predicted that next year's GPT-7 "could directly shatter the core bottlenecks facing embodied intelligence". If even partially correct, the implication reshapes capital allocation across hardware, simulation, and model training.
The big picture
Three distinct developments frame the week. First, Wang's keynote. He said GPT-7 will grant robots "sufficient generalization capabilities and commercial viability to integrate deeply into factories and daily service settings"4. He expects commercial service applications to commercialize faster than industrial use cases, with home scenarios trailing further behind. Unitree is already trialing internally what Wang described as a "self-evolving physical AI system," a self-loop pipeline combining coding, training, deployment, testing, and feedback. Human checks on motion and task performance remain necessary for now, he said, but human involvement will gradually decline with growing data and experience.
Second, BeingBeyond released Being-M0.7, a latent world-action model pretrained on more than 10,000 hours of raw human egocentric video and motion data before it ever sees robot trajectories2. The model works in three stages: pretraining maps video and motion into a shared space, mid-training adapts that prior to the robot's camera view and body, and post-training adds a separate action expert that outputs whole-body commands. Rather than predicting pixels, the model operates on frozen DINO image features, and both humans and robots share a compact 22-dimensional, head-rooted motion representation. The approach addresses a core bottleneck Wang himself flagged: real robot demonstrations for whole-body loco-manipulation are scarce and costly to collect, while human first-person video and motion capture are plentiful.
Third, DeepSeek agreed to lead robotic AI startup Winverse AI's seed funding round of tens of millions of dollars, valuing the startup at roughly $400 million post-money3. ANALYSIS DeepSeek's move marks a foundation-model lab deploying capital directly into robotics hardware and software, not merely licensing models to it.
Meanwhile, UBTECH set up a joint venture with C&D Inc., planning to sell at least 10,000 humanoid robots by 20306. And Shanghai launched a new-generation intelligent manufacturing initiative targeting 8,000 humanoid robots deployed in factories and an intelligent manufacturing equipment industry exceeding 200 billion yuan (approximately $29.9 billion) by 2028.
Between the lines
Wang's optimism carries notable caveats he did not bury. He acknowledged that many robot models "are still in a rudimentary phase, dependent on developers' manual programming, training and feedback, leaving workflow efficiency to be boosted". He also said adding high-performance compute to every robot is "a thankless endeavor," arguing that AI focused on motion capabilities can be deployed on the robot itself while decision-making tasks and large-scale scenario deployments still require cloud computing. ANALYSIS That split, edge for locomotion and cloud for cognition, implies robotics companies remain dependent on cloud infrastructure providers and, by extension, on whoever controls the frontier reasoning models.
Wang Jian, an academician of the Chinese Academy of Engineering, called on manufacturing enterprises to open up their industrial data, saying it is locked in the "drawers" of individual enterprises. He named Haier and Schneider Electric as manufacturers he hoped would take the lead. ◆ The plea suggests that even as foundation models improve, the data pipeline for industrial robotics remains fragmented, a structural bottleneck that model scale alone does not solve.
In a rare moment of reflection, Wang Xingxing admitted his biggest mistake over the past decade-plus was "not believing in AI strongly enough". ◆ The confession doubles as positioning: it frames Unitree's current all-in bet on foundation models as a correction, not a gamble.
What's next
Shanghai's 2028 target of 8,000 factory-deployed humanoid robots provides a concrete benchmark against which Wang's one-to-two-year tipping-point claim can be measured. UBTECH's 10,000-unit sales goal by 2030 sets another. Wang said Unitree had launched pilot deployments across multiple sites including auto factories back in 2024. ◆ The next scheduled test: whether GPT-7, whenever it ships, delivers the generalization leap Wang is banking on.