ModelBest's Zou Zhensheng: Embodied AI Scaling Hinges on Reliability and Customer Economics
At a Sept. 23 seminar in Beijing, the Digital Economy Research Center of China Economic Information Service hosted the "Xinhua Future Opportunity Dialogue: Breaking Through—From Demo-Level Intelligence to Scaled Application" high-quality development seminar on embodied AI. Zou Zhensheng, engineering vice president of Beijing ModelBest Intelligent Technology Co., Ltd., told attendees that embodied intelligence is not a one-company game but an ecosystem collaboration. Only by deepening the general model foundation can scenario transfer become engineering-meaningful. He stressed that the final mile of scaling is not flashy demos but reliability and making customers able to justify the economics.
He argued that the last mile of scaling embodied intelligence is not cool demos but reliability and economic viability for clients. That view framed his later discussion of data, scenarios and commercialization. He suggested the industry must move beyond presentation-level intelligence to large-scale application, where value is proven in operations rather than in showcases. He framed the challenge as moving from demonstrations that impress to systems that work reliably at scale.
On data-mix strategy, Zou said ModelBest was founded in 2022 and grew out of technology transfer from Tsinghua University's Natural Language Processing Lab. Different training stages require different data-mix strategies, a key guarantee for continuous model evolution. He explained that the right mix changes as training progresses, and that this discipline underpins long-term model improvement. He described the data strategy as a moving target rather than a fixed recipe.
He noted that human first-person-view data is currently the only embodied data supply that can scale to tens of millions of hours, giving it the greatest potential to serve as pretraining material for an embodied Scaling Law. But it lacks action labels and proprioception, requiring cross-embodiment alignment. Real-robot teleoperation data is currently the mainstay for post-training and offers the most direct success-rate gains on contact-intensive tasks, though collection costs are high. That combination defines the data challenge for scaling. The gap between available data and required labels remains a central engineering problem.
On the scenario ladder, Zou said the pace of embodied AI deployment depends on three dimensions: task definability—whether success criteria are clear; tolerance for failure costs; and difficulty of data feedback. These yield a clear sequence for commercialization. He grouped deployment scenarios into tiers based on those factors. This framework helps prioritize where embodied AI can deliver value first.
The first tier includes automotive and logistics sorting, where tasks are highly standardized and data feedback is smooth. The second tier is special scenarios, the third is commercial services, and the fourth is home scenarios. Home scenarios are highly non-standard, making task success hard to define uniformly and data feedback most difficult. That difficulty explains why home use is expected to lag other markets. By contrast, home environments demand broad generalization and tolerate little failure.
Discussing core constraints on downstream large-scale commercial use, he acknowledged that the real bottleneck is not only model capability but also reliability and systems engineering. The key is whether customers can make the economic math work. Reliability, not just intelligence, determines scalability, he said. Customers, he said, need clear returns before committing to deployment.
For the next two to three years, Zou predicted the edge large-model industry will unfold along a closed loop of perception, thinking and action capabilities. Language models are the most mature, multimodal models are next, and embodied VLA models are still evolving. The industry landscape will develop step by step rather than all at once. Each capability layer builds on the previous one.
In applications, automotive will scale first. Driven by rising model capability density and stronger edge-chip compute, consumer electronics will follow, and robots will become the largest commercialization scenario. Automotive is expected to lead, with consumer devices next and robotics offering the largest long-term market. That sequence reflects both technical readiness and market demand.
Zou called for embodied intelligence to be an ecosystem collaboration rather than a one-company game. He hoped more industry partners will join to build a multimodal collaborative data loop and evaluation system, while establishing publicly available dataset standards and trading mechanisms to break data silos. He also urged broader collaboration on shared evaluation and open data exchange. Such mechanisms, he said, are essential for sustainable progress.