Expert Talk Robotics
At the CIIF in Shanghai on October 12, Wang Xingxing, founder of Yushu Technology, delivered a highly impactful self-review and industry forecast. Almost simultaneously, two investment moves by DeepSeek emerged: one bet on the robot "body," Yushu, and another planned bet on the underlying embodied "brain," Winverse AI. Large model developers and robot complete machine manufacturers are moving from two parallel lines towards a mutual confluence.
01. Wang Xingxing's Cognitive Reversal: The Biggest Mistake Was Not Believing Enough in AI
The two speeches at the CIIF represent a significant forward revision by Wang Xingxing regarding the expected inflection point of embodied AI.
Previously, his judgment on the industry's critical point was at least three to five years away; however, at this year's CIIF, he directly compressed the optimistic expectation to as fast as one to two years, even proposing a personal prediction: with the release of GPT-7 next year, it is possible to directly break through the core challenges of current embodied AI. He frankly admitted that over the past decade of making robots, his biggest mistake was not believing enough in the potential of AI.
Supporting this judgment is the actual machine testing progress of the GPT-6 series. Yushu has deployed large models onto robots to complete multiple rounds of real-machine tests, and the task generalization ability in unfamiliar scenarios exceeded expectations. However, Wang Xingxing remains clear-headed at the engineering level, proposing a clear implementation route: minimize computing power on the physical body, move decision-making to the cloud, and treat real-machine physical interaction data as irreplaceable.
In his framework, future humanoid robots will not cram a full set of large models into the chassis. The physical body will only retain the basic computing power for motion control, while complex reasoning and task planning will be handed over to cloud-based large models. The core value of the robot is to serve as a collection terminal for the physical world, colliding, grasping, and trial-and-error in real environments, generating physical interaction data that simulators cannot replicate, and moving towards "physical AI self-evolution." In terms of commercialization pace, industrial scenarios and B2B businesses will be implemented first, while home robots will require a longer cycle of refinement.
02. DeepSeek's Two Moves: A Bet on the Body, A Bet on the Brain
In the very same week that Wang Xingxing updated his industry judgment at the CIIF, DeepSeek's embodied chess game was fully revealed.
In August, DeepSeek participated in Yushu's IPO strategic placement with CNY 141 million, with a lock-up period of 36 months. This investment comes with an unprecedented bilateral agreement in the industry: priority procurement of each other's products and services under equal conditions.
When DeepSeek explores embodied business, it will prioritize Yushu's complete machine and hardware solutions; when Yushu undertakes large model adaptation and embodied development, it will prioritize DeepSeek's models and computing power services under equal conditions.
This is not a simple financial investment, but a mutually bound industrial alliance. For Yushu, there is no need to develop large models from scratch; by leveraging mature upstream large models, it can reduce a massive amount of repetitive R&D investment and focus resources on the physical body, motion control, and real-machine data collection. For DeepSeek, simply iterating models in virtual space has a ceiling. Taking a stake in Yushu is equivalent to securing a carrier for its own large models to enter the physical world. Through real-machine interaction data, it can make up for the physical common sense missing in large models.
After positioning in Yushu, news of another heavy investment by DeepSeek emerged: it plans to lead the investment in Winverse AI, founded by Hu Han, co-author of the Swin Transformer, with a post-investment valuation of approximately USD 400 million, and Meituan Longzhu participating in the follow-on investment. Sources familiar with the matter at Bloomberg indicate that this deal has not yet been finalized.
By investing in both the robot physical body and the underlying embodied model team, DeepSeek will complete the dual-line positioning of "body + brain." At the same time, Liang Wenfeng, founder of DeepSeek, is holding intensive meetings with embodied AI entrepreneurs, accelerating the exploration of paths for models to land in the physical world beyond large language models.
03. Thoughts on the Ultimate Industry Outcome: Restructuring of Division of Labor
The core of this round of industrial change is the restructuring of the division of labor.
In the past, many complete machine manufacturers attempted full-stack in-house development, trying to build both the physical body hardware and investing heavily in training large models. This resulted in high R&D costs and long cycles, making it difficult to achieve top-tier results in both areas simultaneously. The route chosen by Yushu is to hand over the cognitive brain to professional large model developers, focusing on leveraging its strengths in physical body hardware, real-machine scenarios, and physical data collection.
On the other side, large models are only trained in virtual worlds such as text and images, and will never learn the constraints of the physical world: object weight, friction, collision feedback, and grasping fault tolerance. To move towards AGI, they must enter the physical world, and the robot physical body is the carrier for AI to acquire physical experience. DeepSeek's investment in Yushu and its layout in Winverse AI are essentially about making up for the physical world lesson for large models.
Conclusion
Wang Xingxing's remark at the CIIF, "My biggest mistake was not believing enough in AI," and DeepSeek's investment moves grasping both "body + brain," constitute a footnote to the same era.
The inflection point of embodied AI is rapidly approaching, and hardware manufacturers and AI companies are no longer fighting alone. The two are moving towards each other, forming a brand-new paradigm: the hardware physical body provides the physical body and real interaction data, while the large model provides the cognitive planning brain, and the two form a closed-loop iteration.
This article is compiled based on publicly available network information and is for reference only, not constituting investment advice.