The boundary between foundation models and embodied intelligence is blurring. In Liang Wenfeng's R&D roadmap, embodied intelligence is placed at the very end, yet DeepSeek's investment layout has already taken the lead.
DeepSeek previously invested in Unitree Robotics, and now plans to lead the investment in Winverse AI, founded by Hu Han, the former head of multimodal understanding at Tencent Hunyuan. One excels in robotics hardware and motion control, while the other explores embodied intelligence. DeepSeek's R&D focus remains on foundation models, but its investments have already reached into the physical world.
These two moves are not contradictory. In the route envisioned by Liang Wenfeng, it is a better sequence to first endow AI with the ability of continuous learning and self-iteration, and then leverage a stronger AI to develop robots. The more proficient AI becomes in programming and reasoning, and the more work it can take on during the R&D process, the less likely it is to take detours when entering the embodied field. Conversely, if these capabilities are not yet sufficient, more effort must be spent on robotics R&D, tackling problems one by one.
Therefore, embodied R&D can be done later, but the layout can be advanced first. DeepSeek continues to concentrate its efforts on advancing foundation models, while through investments, it gains early access to robotics hardware, research talents, and development and testing resources. When it truly enters the fray, it may not need to catch up from scratch.
For embodied AI vendors, this is what truly warrants vigilance. Future opponents may be cross-over players who have already accumulated stronger capabilities in foundation models and made preparations in advance. Today, it is still an investor and potential partner; tomorrow, it may compete with you for the core intelligence of robots.
01. Focusing on General AI While Advancing Strategic Layout
To bring AI into the physical world, embodied intelligence requires more than just models. Whether a model can understand the immediate environment and provide appropriate actions must ultimately be tested on a robot. What Unitree can provide is precisely the foundation for carrying out this work.
People are familiar with Unitree's robot dogs and humanoid robots, but its products also include robotic arms, dexterous hands, and key components such as joints and motors, along with corresponding software and development interfaces. For model teams, this means they can start testing their ideas without having to build a robot from scratch first.
Testing, however, brings new problems. When the robotic arm reaches out, will it bump into things? If the grasping position is correct, why is it still not holding steadily? Some of these issues stem from the model, while others arise from control and hardware, requiring repeated joint debugging.
Unitree's years of accumulated robotics R&D experience hold value here. Investing in Unitree provides DeepSeek with an entry point to access this hardware, tools, and development experience.
Winverse, on the other hand, connects to another type of accumulation: how to enable models to understand the physical world and learn to act.
Founder Hu Han also has research ties with DeepSeek. He and DeepSeek researcher Xie Zhenda previously participated together in research on SimMIM, Swin Transformer V2, and others, collaborating on visual models and large-scale training. Therefore, this proposed investment is not only about engaging with an embodied intelligence team but also continuing existing research connections.
Through these two investments, the hardware, talents, and cooperative relationships are prepared first. DeepSeek keeps its main R&D forces on foundation models while retaining the opportunity to delve deeper into the embodied field. The advantage of this arrangement is that Liang Wenfeng does not need to bet on the outcome of the two routes right now. If the progress of foundation models can drive robotics R&D, DeepSeek has the opportunity to accelerate its entry through these collaborations; if embodied intelligence still requires long-term independent exploration, it is already involved. Focusing on foundation models and advancing the layout for embodied intelligence can proceed simultaneously.
02. From Action Decision-Making to the Competition in Training Systems
Liang Wenfeng's judgment is not isolated; Dario Amodei, CEO of Anthropic, has had a similar vision.
In "Machines of Loving Grace," Amodei depicts a powerful AI that may not have its own body but can operate robots and experimental equipment through computers, and even design robots for its own use. The two judgments share an important intersection: the stronger the foundation model, the more likely it is not only to manipulate robots but also to help people develop them.
Google has already advanced in this direction. Its Gemini Robotics system has corresponding layouts ranging from understanding the physical environment and generating actions to running models locally on robots. This indicates that general model companies started seriously addressing the problem of how robots act relatively early.
The performance of GPT-6 Astra, recently released by OpenAI, makes this cross-over more intuitive. In a study by RoboDojo, researchers provided it with images, robot states, and task instructions, allowing it to decide where the robotic arm should move next, what orientation to adjust to, and how the gripper should act, before handing it over to the control software for execution.
Astra was not specifically fine-tuned for these evaluation tasks, yet it achieved an average success rate of 22.48% across 42 simulation tasks and a total of 2,100 trials; for reference, two robot policy models in the same study, DM0.5 and π0.5, achieved 19.34% and 6.91%, respectively. The input conditions were not entirely identical, but the results are already sufficient to make embodied AI vendors take it seriously: foundation models are beginning to show competitiveness in robot action decision-making.
Deciding how the robotic arm should move next was originally the core problem that embodied models needed to solve. Now, foundation models can also participate in the competition.
Where does this capability come from? A common misconception is that Astra originally only handled tasks in the digital world and later suddenly learned to operate robots. In fact, the lack of specific training during evaluation does not mean it hasn't learned related content before; industry insiders generally believe that Astra's training utilized a massive amount of robotics data.
Embodied AI vendor Physical Intelligence (PI) conducted an experiment that helps clarify this relationship. When training π0.5, it used both robot action data and image-text data from the internet. In a test requiring the robot to manipulate unfamiliar objects, the average success rate of subtasks with the complete training combination was 94%; after removing the internet image-text data, the score dropped to 74%.
Teaching robots to work does not necessarily rely solely on repeatedly demonstrating actions. Image-text knowledge from the internet can also help them better understand the objects in front of them and the tasks assigned by humans, thereby performing the work well.
This brings up a new competitive issue. In the past, people cared more about who collected more embodied data and who did a better job in action training. Now, it also depends on who can combine this data with broader knowledge. The capabilities accumulated by foundation model vendors over the years may come in handy right here. Embodied AI vendors can certainly adopt this method as well, but what they need to face is already a competition of a different scale and approach.
There is another path, which aligns more closely with the visions of Liang Wenfeng and Amodei: letting AI help develop robots.
In another experiment of the aforementioned study, Astra wrote a program to control a dexterous hand to play the piano, and then modified the code based on simulation results, participating in the design and debugging of the control scheme. At this point, it was taking on a portion of the research and development work.
The earlier Eureka study also involved GPT-4 in designing robot training rules, telling the training system which actions were performed well and which should be avoided. The subsequent DrEureka went a step further, allowing language models to design simulation training conditions, helping robots apply the skills learned in simulation to the real world.
These attempts cannot yet let AI take over robot R&D entirely, but they already demonstrate that capabilities in programming, reasoning, and tool invocation can enter the robot development process. Latecomers, leveraging stronger AI, may write control programs and design training schemes faster, continuously discovering and adjusting solutions in simulation and testing. The strategic significance of Liang Wenfeng prioritizing foundation models also lies here: first improve AI's ability to solve R&D problems, and then use it to lower the difficulty of entering the embodied field. The experience accumulated by pioneers over many years might be easily pierced by latecomers.
03. Embodied AI Vendors Must Embrace Challenges at the Foundation Model Level
Of course, the fact that AI can help write programs and design training schemes does not mean that robots can perform tasks well in reality.
Wang Qian, founder of X Square Robot, reminds that foundation model vendors entering the embodied field must also solve problems such as data collection, simulation, evaluation, and hardware. How to adjust force when a robot touches something, how to react when an object suddenly slips, and whether it can maintain stability after working continuously for hours all require repeated exploration in real environments.
Astra also faces a very practical problem: it is slow.
NVIDIA's Astranomy research tested Astra in a simulation environment, including the waiting time when the model generates actions in the testing process. In the mode of directly generating actions, each inference call of Astra takes an average of about 18.56 seconds, which includes network latency.
The simulation environment can pause and wait for the model to return results, but the real world will not. A robot reaching out to catch a slipping cup does not have ten-odd seconds left for it to think.
Therefore, there is still a gap between being able to complete tasks in experiments and completing them timely and reliably in reality. This gap is precisely where the years of accumulation by embodied AI vendors can play a role.
Embodied AI vendors are not sitting idle waiting for foundation models to catch up. Generalist chooses to train its own foundation models from physical interactions, hoping that robots can learn more general capabilities in the process of contacting and manipulating objects. Deng Taihua, Chairman of AgiBot, states the route very directly: "Embodiment first, then intelligence, not intelligence first, then embodiment."
This sequence differs from Liang Wenfeng's choice. Liang Wenfeng hopes to first strengthen the foundation models and then use them to reduce the difficulty of robotics R&D; AgiBot emphasizes more on letting robots enter actual environments first, accumulating experience through use, and then promoting intelligence enhancement. Both parties want to build stronger models, just starting from different places.
Therefore, simply viewing the entry of foundation models as a comprehensive "dimensionality reduction attack" underestimates what embodied AI vendors are doing. They are also building foundation models and can utilize them to improve training and R&D. Physical experience is the advantage of embodied AI vendors, but entering early does not mean they can hold their ground. The key is whether they can use accumulated experience to improve models, enabling robots to collect less data and do less debugging when learning new tasks. Only by achieving this can they have the confidence to compete against foundation model vendors.
04. Holding onto Core Intelligence, or Yielding More Value
Figure has already made a choice: ending its collaboration with OpenAI and turning to in-house development.
In an interview at the end of April 2026, Figure founder Brett Adcock explained that robot training, debugging, repair, and quality control need to be promoted together in coordination. If the core model is held by an external company, it is difficult to improve these work together according to their own needs.
Ultimately, Figure does not want the core capabilities of its robots to be constrained by someone else's model.
But deciding to do it in-house is just the beginning. If a robot has to recollect a batch of data every time it learns a new task, and then engineers repeatedly debug it, then even if the model is developed in-house, it may still be caught up by latecomers. The true moat is turning today's encountered problems and accumulated experience into the ability to learn new tasks faster tomorrow.
Embodied AI vendors can also utilize foundation models to accelerate this process. AI can help foundation model vendors write programs, run simulations, and analyze test results, and it can equally help embodied AI vendors do these things. Existing physical experience, combined with increasingly powerful AI tools, may also allow pioneers to run faster.
Therefore, this competition is more likely to lead to the differentiation of embodied AI vendors.
Some vendors will gain a foothold in foundation models, competing with foundation model vendors for the core intelligence of robots. The strengths of other vendors may remain in hardware, system integration, and actual delivery. They can still become excellent robotics companies, but who manufactures the robots and who endows them with intelligence may no longer be the same answer. The most valuable parts may not all remain in the hands of embodied AI vendors.
If they can neither build competitive models nor have obvious advantages in hardware and delivery, their days will be tough. With foundation model vendors entering from the front and stronger embodied peers beside them, the advantages brought by early financing and early entry will be hard to sustain.
Looking back at DeepSeek's layout, Liang Wenfeng has left himself with more choices. Internally, it continues to focus on foundation models; externally, it gains early access to embodied talents, hardware, and development and testing resources. Whether foundation models gradually learn physical tasks or the training of foundation models and embodied models further integrates, DeepSeek has the opportunity to participate.
This is Liang Wenfeng's overt strategy in embodied AI: R&D can have a sequence, but the layout does not need to wait until the end. First strengthen the foundation models, and then use stronger AI to enter the physical world; at the same time, prepare the potentially needed partners and conditions in advance. For embodied AI vendors, such an opponent is harder to deal with. It can cooperate with you, and it may also incorporate the tasks you are tackling into its own model training. When it truly enters the fray, what it brings may no longer be just capital, but also a new set of problem-solving methods.