Reports Detail DeepSeek's Moves on Embodied Intelligence

On Oct. 13, the 21st Century Business Herald, citing a report by Phoenix Net Technology, reported that Liang Wenfeng was involved in discussions related to embodied intelligence at DeepSeek. According to the report, no specific plan has been decided, and the company is weighing two options: building its own team or working with investment-related companies. As background, Liang reportedly said at a closed-door exchange in July that world models have "little to do with the upper limit of intelligence" and that DeepSeek would "not do it." He did not rule out embodied intelligence, however, and reportedly placed it along the extension of the company's road map from this year onward.

If Liang's reported comments are accurate, DeepSeek has taken the position that world models are outside its core strategy, making any move into embodied intelligence a measure of how far it is changing course. The report points to more than an interest in a new topic: it raises the question of how much a model company will invest in data from the physical world and control of real-world machines.

What appears new in this report is not whether DeepSeek will enter embodied intelligence, but which route it might take: building its own team or investing. Liang reportedly said at a closed-door exchange in July that 3D generation, video generation and world models were outside the company's core strategy, and that multimodality would remain a "component." The latest reports can therefore be viewed as a possible adjustment along that trajectory. However, the report says the policy is undecided, and it is not yet clear whether the company will move toward commercialization. This development lies between two related reports by this newspaper: Artificial Analysis unveils AA-Robotics v0.1, with GPT-6 Astra in first place and [ABB, FANUC and others show robots "working in factories" at the industrial fair, putting the race to deploy embodied intelligence in the spotlight](/news/kouhaku-kai-de-abb-ya-hatsu-na-ka-raga-koujou-de-hataraku-te-0c047f). The first showed how benchmarks for real-world operations could shift from text processing to robot control; the second reported that factory applications are coming to the fore. If DeepSeek's involvement is confirmed, the decision would be whether to enter a competition not just among models, but over how computing resources are linked with physical-world data. Publicly available information shows that competition in embodied intelligence cannot be confined to software alone. The learning cycle depends on connecting robot control, world models, simulation and real-world data, so where DeepSeek allocates resources will significantly shape the meaning of any entry. Liang is also reported to prioritize finite resources. Without clarity on how embodied intelligence would fit with the company's existing inference models and agent strategy, it could remain a peripheral investment. As ladderobot.com reports on differences between embodied intelligence in China and the United States has shown, Chinese companies are often described as racing to scale up. If DeepSeek were to enter the field, the key questions would be whether it would focus on research, build an ecosystem through investment, or take on data collection and real-world operations itself. Three points warrant watching. First, whether the company opts to build its own team or work with investment-related companies. Second, how it distinguishes embodied intelligence from world models and agents. Third, how much of its real-world data and simulation infrastructure it builds in-house.

The direct implication for Japanese companies is that embodied intelligence depends not only on model performance but also on the combination of real-world data, simulation and robot control. DeepSeek's specific plans remain undecided based on publicly available information. If Chinese foundation-model companies enter the field, the competitiveness of Japan's robotics, component and testing-environment suppliers could become a prerequisite for success.