Embodied AI Leaders Clash Over Data, Models and Commercialization at Bund Summit

The first divide concerns data. Han Zheng said real-world data collection has limited efficiency and unstable quality, making simulation an unavoidable part of the stack. Sudo Technology has long bet on simulation, hoping large-scale reinforcement learning will give robots generalization across different objects and environments. He stressed that simulation and real data are not mutually exclusive and will more likely be layered and combined.

Shen Yujun pushed the question to the physical world. Once robots enter real environments, all information comes from onboard cameras, tactile and force sensors. Real sensors have noise, errors and individual differences, and these imperfections are part of reality. Therefore, more important than 'Sim to Real' may be 'Real to Sim': whether complex, fine-grained physical features from the real world can be moved into simulators at scale. Tactile sensors, for example, produce real signals with complex variations in frequency, amplitude and consistency that are not easily replicated. The valuable data competition may not be who has more hours of data, but who can obtain and organize experiences that truly contain physical intelligence.

Xu Huazhe warned that the industry overestimates data volume. 'One million hours' or 'several million hours' does not automatically mean a better model. Beyond scale, quality, distribution and training methods may matter more.

The second divide concerns intelligence. Recent performance of GPT-6 Astra in robot control renewed debate over whether general large models will enter robots directly and 'crush' embodied AI companies. Wang Qian disagreed. His core view is that 'the ontology of intelligence exists in data.' Models are more like distillers in training and containers in deployment. Coding advanced quickly not because language models naturally acquired programming ability, but because code data is easier to generate, verify and feed back. If a general model is to enter embodied AI, it still must confront robot data, real-world evaluation, hardware, simulation and deployment infrastructure. These difficulties will not disappear just because of a stronger language model.

Xu Huazhe also tested Astra and observed strong semantic understanding and 3D spatial ability. It could identify objects, grasp chopsticks and even complete some delivery operations. But in contact-rich tasks such as folding clothes, physical interaction remained clearly insufficient. The more likely future is not replacement but layering: upper-level large models handle semantics, reasoning, planning and task decomposition, while lower-level embodied models handle real-time perception, continuous control and reliable execution. Shen Yujun said this layering may in turn change large models. Robots must continuously receive sensor input during action and change decisions in real time; they cannot operate like today's chat models in a question-and-answer mode. This means embodied AI may eventually form a training paradigm of its own for the physical world. 'Embodiment may follow the same model as the digital world, but it will definitely not follow the same technical route,' Shen said.

The third divide concerns scenarios. If data and models determine the technical ceiling, commercialization determines whether the industry can survive cycles. The panelists gave different yardsticks for what counts as entering the industrial stage, but all pointed to whether robots can create real value stably and at scale. Han Zheng emphasized success rate and generalization. In real commercial settings, robots and lab demos face completely different requirements. Many industrial tasks must achieve close to 100% success on first execution, while also adapting to different objects and environments, rather than relying on extensive post-training to 'overfit' each new scene.

Shen Yujun said the industry can fall into another trap: attributing commercialization almost entirely to model performance. In real scenarios, customers buy a working system, not a model metric. Will a full humanoid robot be used, or only a dual-arm setup? What is the hardware cost? Are sensors reliable? Can the system run stably for a long time? Can deployment and maintenance costs be accepted by customers? These questions directly determine whether commercialization holds. 'Embodiment is complex systems engineering,' Shen said. In his view, the model is only one link. Real deployment requires considering hardware, models, system robustness, deployment methods and final ROI, rather than reducing all scenario problems to 'the model is not strong enough.'

Wang Qian offered a more direct standard: whether customers keep paying and pay more. In his view, the real commercial value of this generation of robots should not be merely replicating the previous generation of automation. It should solve problems that could not be solved before and produce better ROI than manual labor or traditional automation in real scenarios. Once embodied AI companies can continuously create customer value through frontier technology, commercialization revenue will in turn support larger models and more R&D investment, forming a flywheel of technology and business.

Xu Huazhe said the true leap for this generation of embodied intelligence is for foundation models to keep improving success rates on large numbers of unknown tasks, from 50%, 60% and 70% toward the 99.99% required in industrial settings. The answer to the scenario debate may not be finding a single killer app, but crossing several thresholds at once: models must generalize, execution must be reliable enough, system costs must work, and customers must be willing to keep paying.

For the next five years, the panelists were generally more optimistic. Shen Yujun said it is entirely possible that a completely new technology will emerge. Xu Huazhe judged that general embodied services may not need five or ten years: 'Within three years, it can begin serving in many places.' Wang Qian said people may even underestimate the importance of embodied intelligence itself. Agents in the digital world have already begun restructuring software production, and only when AI truly gains the ability to transform the physical world can AI's transformation of production systems form a complete loop. Looking back five years from now, the most important thing may not be which demo was most impressive, but which technical route first turns physical intelligence into stable, replicable and scalable productivity.

The 2026 Inclusion Bund Summit was held under the theme 'Co-creating the AI New Economy.' It featured one main forum and more than 40 insight forums, bringing together more than 600 scholars, scientists, entrepreneurs, investors and innovators from China and abroad. The summit also held a technology exhibition, developer day and venture capital meetup to connect technology research, industrial application and innovation.