Executive Summary
At the July 2026 World Artificial Intelligence Conference (WAIC) in Shanghai, phrases like "the year of the world model" and "the GPT moment for embodied intelligence" circulated across dozens of forums. But inside several sessions, a different current ran through the program. University College London (UCL) professor Wang Jun told an audience that most embodied reinforcement-learning claims are not actually trained on real robots but rely on distillation, stating flatly that "theoretically, it can be proven that if you use this distillation method, you cannot generalize." Stephen Redmond, a professor at University College Dublin, said he does not believe vendor claims of "zero-shot generalization." And Elite Robotics chairman and CEO Cao Yunan told an industry forum, on the record, that the sector "has to admit it's only at the L2 level, not even L3," with no independent third party verifying the 99% success rates companies routinely cite. These voices form an important counterweight to WAIC 2026's dominant embodied-AI narrative.
Industry Context
Two phrases dominated embodied-AI discourse at WAIC 2026: world models and the GPT moment. Fang Han, chairman and CEO of Kunlun Tech, told his company's forum, "We firmly judge that 2026 is Year One of the world model." SenseTime chairman and CEO Xu Li, speaking at the "World Model Six Dragons" summit forum, said, "In history we witnessed the ChatGPT moment; today we are witnessing another equally important inflection point," citing a Morgan Stanley forecast that global humanoid robot shipments will grow from 50,000 units in 2026 to 445,000 units by 2030, a 106% compound annual growth rate for 2025-2030.
Against that backdrop, a group of academicians, senior researchers at multinational R&D operations, and industry veterans chose public forums to push back, concentrating their critiques on three areas: whether generalization claims are overstated, whether model opacity is being underweighted as a deployment risk, and what stage of maturity the industry has actually reached. None of these critiques reject the embodied-AI technology path outright; rather, they demand a stricter evidentiary standard for claims already circulating in the market.
Technology: Methodological Challenges to Generalization, Distillation, and Benchmarks
During the roundtable at forum xa12 ("Frontiers and Industrial Ecosystem of Embodied Intelligence" at WAIC 2026), Stephen Redmond, professor and head of the biomedical sensors and signals research group at University College Dublin, said plainly: "If they just mention they can do zero-shot generalization, I usually don't believe such claims." He argued that evaluating generalization requires specifying exactly which factor — geometry, appearance, lighting — actually transferred; without that specification, "zero-shot generalization" functions as an unfalsifiable marketing term. He also noted that establishing benchmarks in robotics is far harder than in computer vision: other labs struggle to replicate test conditions and cannot verify whether an experiment was correctly implemented, and he suggested independent third-party testing facilities — citing the Robot Cup model — as a possible way forward.
At the same roundtable, Beihang University professor Jiao (full name unclear in the transcript) used an early-20th-century chimpanzee experiment as an analogy: a chimp that had never seen a particular environment could still use a stick to retrieve a banana from the ceiling, because it possessed a reusable, composable skill rather than data covering every possible scenario. He argued current models lack exactly this: "I don't think the Transformer will be the endpoint of embodied or general intelligence either," because language is discretely compositional while robots operate in continuous physical environments. He, too, said he distrusts papers and companies that claim zero-shot transfer without specifying which generalization factor was actually tested.
At forum xa28, hosted by the Hong Kong University of Science and Technology, UCL professor Wang Jun delivered the sharpest methodological critique of the conference. He argued that foundation models have reached a point of data exhaustion, and that the core scientific challenge in moving from the digital to the physical world is continual learning: "How could it possibly learn within a few rounds of iteration in the physical world? This is almost Mission Impossible — it calls for new algorithms." More pointedly, he stated: "Theoretically, it can be proven that if you use this distillation method, you cannot generalize" — describing a widespread industry practice in which companies claim reinforcement learning but in fact train through distillation rather than on real hardware. He identified learning by interaction — acting in an environment, receiving feedback, and improving — as the real unsolved problem, compounded by the difficulty of even defining reward functions in the physical world.
Engineering Analysis: Compounding Error, Black-Box Models, and "Hands Designed for Machines"
These methodological concerns translate into concrete, quantifiable engineering obstacles.
The first is error accumulation over long-horizon rollouts. At forum xa25 (Kunlun Tech's World Models and Multimodal Paradigm Shift forum), Zhou Zhihua, a Chinese Academy of Sciences academician and vice president of Nanjing University, cited an earlier proof by Michael Kearns showing that multi-step rollout error compounds quadratically with the number of steps: "You start from like 100 steps, right? There will be an error of like 10,000 times." This is why the academic community long considered decision-layer world models a non-starter. At forum xa05 (Physical AI Reshaping Intelligent Manufacturing), Nanjing University AI school professor Yu Yang independently cited nearly the same figure, concluding it is unlikely there will be "one giant" world model, since industrial scenarios demand precision far beyond what language models require, where a human can catch mistakes. Two academicians arriving independently at the same order-of-magnitude warning suggests this is not an isolated observation but a structural bottleneck in applying current world-model techniques to decision and planning tasks.
The second is model opacity. Li Zhijun, founder of Shanghai Juishi Intelligent Technology and a Cheung Kong Scholar professor at Harbin Institute of Technology, told the xa12 roundtable that the deepest pitfall in post-training is that the model is uninterpretable: "Every time I watch the scene I want to step up and say a few words — 'robot, you should do it this way' — but it doesn't listen; it can't understand." He argued this black-box property forces deployment teams to rely purely on feeding more data, while ordinary factory customers have no idea how to collect data effectively. He also offered a design provocation worth engineering attention: current dexterous-hand development largely mimics the human hand, but "the design of the human hand is for humans... for the dexterous hand, I think they should be designed for machines." In B2B industrial settings, he said, advanced dexterous hands are "both expensive and unsuitable" — a more realistic near-term path may be quick-swap, modular tools built for robots rather than five-fingered replicas of human hands.
The third is power and energy constraints. Guo Yandong, founder and CEO of AgiBot rival AI2Robotics (智平方), offered what he called a "somewhat blunt" claim at forum xa28: "I think a major limiting factor preventing the embodied large model from entering this GPT moment might be power supply," noting that training an embodied model may require many times the electricity of training a language model. Wolfram Burgard, founding chair of computer science and AI at the Technical University of Nuremberg, raised a related infrastructure concern at forum xa12: "In the US, they are basically speaking about building large compute centers, facilities together with a nuclear plant next to it" — a scale of infrastructure investment he said is very difficult to replicate in Europe, calling it a major challenge for the region's embodied-AI development.
Commercial Progress: Candid Maturity Assessments and the Missing Third-Party Standard
If academic critiques centered on methodology, industry executives directed their skepticism at the honesty of the commercial narrative itself.
At the closing session of forum ts13 ("Toward Autonomy: Real Challenges and Solutions for Industrial AI Deployment"), Elite Robotics chairman and CEO Cao Yunan mapped physical-AI paradigm evolution onto the autonomous-driving L-level framework and stated directly: "We have to admit it's only at the L2 level, not even L3." He described a "negative flywheel": insufficient physical data leads to models lacking generalization, which leads to frequent execution failures, while success-rate claims lack any independent standard — "Every company now claims 99% success in some factory, some scenario... Really? Who evaluates that? Is there any genuine third-party fair standard? No."
Mech-Mind Robotics founder and CEO Shao Tianlan offered a blunter piece of advice at the same forum: "Don't believe anything you see online about robots. Offline demonstrations are far more convincing." He proposed a nine-level product maturity model, noting that many demos remain at levels 0-3, while a company needs to reach at least level 8 to have a legitimate revenue business, and level 9 to be profitable. He pointed out that global annual shipments of industrial and collaborative robots total roughly 500,000 units against approximately one billion manufacturing and logistics workers worldwide — a gap he attributed not to a lack of technology concepts but to the high cost, long cycles, and lack of flexibility inherent in traditional custom-integrated automation. Customers want fast deployment, minimal customization, and reliability — not the term "embodied intelligence" itself.
Notably, the industry has begun responding to these critiques with institutional mechanisms. At forum sb77 (the World Model "Six Dragons" summit), Li Hongsheng, associate dean of Shenzhen Hetao College and secretary-general of the Physical IQ platform, said existing benchmark sets are mostly built by single companies, creating "the awkward situation of being both referee and athlete," which undermines industry-wide trust in evaluation results. The Physical IQ evaluation platform — jointly launched by the Shanghai AI Industry Association, Shenzhen Hetao College, and the East China branch of the China Academy of Information and Communications Technology with dozens of universities and companies — was created to provide unified, third-party evaluation that "resolutely abandons traditional toy tasks divorced from reality." At the same forum, Genie AI (极佳视界) co-founder and chief scientist Zhu Zheng admitted what he called "the most painful lesson": that insufficient model capability, not any commercial or integration factor, is the fundamental reason deployments still fail.
Market Perspective: The Industry Won't Pay for "Smart"
Tao Dacheng, chief scientist at Daxiao Robotics and a Fellow of the Australian Academy of Science, moderating a roundtable at forum sb77, offered a line repeatedly cited throughout the conference: "The industry will not pay for the intelligence; the industry will pay for constant intelligence." He explained that the yardstick for a demo is how good its best run is, while the yardstick for deployment is how bad its worst run is — two fundamentally different measures. This framing offers a more precise lens for the conference's broader arguments over success rates and generalization: the question is not whether a system can complete a task under ideal conditions, but whether its failure modes under the worst conditions are controllable and predictable.
Nobel laureate economist Thomas Sargent of New York University, speaking online at the same forum, brought a cross-disciplinary warning from rational expectations theory. He noted that learning algorithms do not converge to a rational expectations equilibrium but only to a "self-confirming equilibrium": agents are accurate about frequently repeated events but systematically biased about rare, low-frequency ones. Though not stated as a direct critique of robotics, this economic conclusion precisely describes a core risk in current embodied-AI training paradigms — systems that learn from interaction data have a structural blind spot in the long-tail, low-frequency failure scenarios that matter most for safety in industrial and household deployment.
HKUST provost and Chinese Academy of Engineering foreign academician Guo Yike, speaking at forum sb94 (the "Top Thought IP" Thinkers Forum), offered a caution aimed at investors: "Today people rush to invest in physical AI models, world models — lots of buzzwords. But you can't invest by chasing names; you must know where it sits on the map of AI's development." He added that while AI as a whole civilizational shift is unlikely to be a bubble, individual technology trends will rise and fall, and rational investment requires understanding a given technology's actual position within that broader development trajectory rather than being driven by terminology alone.
Challenges
Taken together, the challenges raised by academics and practitioners at this year's WAIC cluster around four interrelated dimensions: generalization lacks a rigorous, verifiable evaluation standard, and terms like "zero-shot generalization" resist falsification absent independent replication; mainstream reinforcement-learning practice frequently relies on distillation rather than real-robot training, and theory suggests distillation carries an inherent generalization ceiling; model opacity prevents engineers from correcting errors on the factory floor and makes it difficult for smaller customers to collect data that would actually improve the model; and the compute and power infrastructure required for training and deployment may prove a harder constraint than the algorithms themselves, particularly in regions with limited energy infrastructure such as Europe.
An observation from forum sb59 (the Shanghai International Port Group's "AI-Enabled Smart Port" forum) adds another dimension. Zheng Qinghua, a Chinese Academy of Engineering academician and Party secretary of Tongji University, noted that current large models still lack a comprehensive understanding of causality, temporal logic, and human factors in the physical world, and continue to face hallucination, black-box behavior, and catastrophic forgetting in extreme scenarios involving coordination between large models and robots. These same limitations, he noted, also appear in the highly polluted, corrosive, confined operating environments of port automation — evidence that they are not isolated defects in a single demo but a recurring pattern across industries.
RobotToday Analysis
Further reading: The methodological critiques documented here pair well with the data-scale-vs-quality and end-to-end-vs-modular debates mapped out in RobotToday's companion framework piece, Physical AI Landscape: From Digital Intelligence to the Embodied Physical World.
Viewed together, the critical voices at WAIC 2026 are not isolated instances of skepticism but converge on a single technical judgment: the biggest uncertainty in embodied AI today is not whether the field can produce impressive demonstrations, but whether it can build a credible, reproducible, third-party-verifiable evaluation framework. Wang Jun's theoretical argument about distillation's generalization ceiling, the matching "10,000x error over 100 steps" figures independently cited by Zhou Zhihua and Yu Yang, and Redmond's questions about benchmark reproducibility together sketch a shared technical consensus: most current progress remains confined to controlled conditions and has not yet crossed the threshold of reliable operation in unknown environments.
The candor at the commercial level is equally significant. Cao Yunan's public admission that the industry is "not even at L3," Zhu Zheng's acknowledgment that insufficient model capability is the fundamental reason deployments fail, and Shao Tianlan's warning not to trust online videos all came from executives with direct commercial stakes in the technology's success — carrying weight comparable to the academic critiques, and suggesting the industry is already generating internal pressure toward self-correction. The launch of the Physical IQ third-party evaluation platform is direct evidence that this pressure is translating into institutional infrastructure.
For engineers, system integrators, and investors evaluating any embodied-AI proposal, the signal from this year's conference is clear: prioritize three questions before accepting a vendor's claims. Exactly which variable was tested when a company claims "generalization"? Was the cited success rate reproduced by a party independent of the vendor? And does the reinforcement learning or post-training the company describes actually happen in simulation, through distillation, or on a real robot? Tao Dacheng's line — that the industry will not pay for intelligence, only for consistent intelligence — may be the most useful summary of this year's sober voices: the next phase of competition in embodied AI is unlikely to be won by the most impressive demo, but by whoever can first give a precise, evidence-backed answer to how bad the worst run really is.
Quotes and figures in this article are drawn from conference public remarks at WAIC 2026 and have not been individually verified with the speakers or their organizations. Where a speaker's name could not be confirmed from the source transcript, this is noted in the text. Corrections are welcome.
WAIC Review 2/4: World Models vs. VLA: Robotics' Next 'ChatGPT Moment'
WAIC Review 3/4: What Scientists and Skeptics Really Think About Embodied AI: The Sober Voices at WAIC 2026
Leave a comment