DeepSeek Eyes Embodied AI: Liang Wenfeng in Frequent Talks with Founders

The embodied AI track is about to welcome a heavyweight potential player. According to Sina Tech, DeepSeek^Ad founder Liang Wenfeng has recently been paying close attention to the embodied AI sector, engaging in discussions with multiple entrepreneurs to explore directions such as robotics, large models, and interaction with the physical world.

According to sources familiar with the matter, Liang Wenfeng is considering the possibility of DeepSeek entering the embodied AI field, potentially by personally assembling a team to develop in-house, or by investing in relevant companies to gain entry.

Investment Landscape and Industry Signals

Liang Wenfeng's positioning in the embodied AI field is not a spur-of-the-moment decision. He had previously invested in Unitree Robotics, and according to foreign media reports, DeepSeek plans to lead an investment in Winverse AI, a robotics company founded by former Tencent AI scientist Hu Han.

This series of moves points to a clear judgment: the integration of large models with the physical world is becoming a question that AI giants must answer.

One detail worth noting: DeepSeek has not responded to this. But judging from the investment path, "invest first, then build" or "invest while building" is the more likely strategy — using capital reach to gauge the industry's temperature, then deciding on the timing and direction for in-house development.

Unitree Robotics was founded in 2016 and is headquartered in Hangzhou. It is one of the leading players in the global quadruped robot (commonly known as "robot dog") space, with its Go and B series products widely adopted in research, inspection, performance, and other scenarios. Since 2023, Unitree has entered the humanoid robot track, launching products such as H1 and G1, and has attracted industry attention with its highly aggressive pricing strategy. Liang Wenfeng's investment in Unitree means that DeepSeek's capital reach has extended to robot hardware — a key link in embodied AI.

Winverse AI was founded by former Tencent AI scientist Hu Han. Hu Han previously conducted computer vision and multimodal research at Tencent Youtu Lab, with an academic background and engineering experience spanning both perception and generation. If the news that DeepSeek is leading an investment in Winverse AI is accurate, it means Liang Wenfeng's strategy is not just betting on robot hardware, but covering both the "brain" (models) and the "body" (hardware), forming a more complete embodied AI puzzle.

Talent War Kicks Off Simultaneously

Echoing DeepSeek's moves, talent flow in the embodied AI field is accelerating.

Former UBTech co-CTO Liao Hongtao has joined Tencent Futian Laboratory as head of embodied AI, and his team has already begun recruiting relevant talent. Futian Laboratory, officially known as the "Guangdong Provincial Laboratory for Embodied AI," is jointly built by the Guangdong-Hong Kong-Macao Greater Bay Area Digital Economy Research Institute (IDEA Research Institute) and Tencent, focusing on embodied AI for human living environments.

Almost simultaneously, Unitree Robotics founder Wang Xingxing admitted during a dialogue session at the Industry Expo: "The biggest mistake I made was not believing in AI enough." He revealed that the team has deployed its latest large model to robots for task testing, with good results, and judged that the critical point for the embodied AI industry may arrive within one to two years.

With big tech companies and startups pushing from both ends simultaneously, the competitive landscape of embodied AI is rapidly taking shape.

Futian Laboratory, officially the "Guangdong Provincial Laboratory for Embodied AI," is jointly built by the Guangdong-Hong Kong-Macao Greater Bay Area Digital Economy Research Institute (IDEA Research Institute) and Tencent, focusing on "embodied AI for human living environments." The so-called "human living environments" refer to spaces highly coupled with daily human life — homes, nursing homes, hospitals, shopping malls, and the like. These scenarios impose requirements on robot interaction safety, generalization capability, and cost control that differ from industrial settings.

Liao Hongtao previously served as co-CTO of UBTech. UBTech is one of China's earliest companies engaged in humanoid robot R&D, listed on the Hong Kong Stock Exchange at the end of 2023 as the "first humanoid robot stock." Liao Hongtao's move from a startup to a big-tech-led laboratory reflects the current trend of embodied AI talent flowing back from hardware manufacturers to big tech companies with large model and computing power resources — a trend that could reshape the industry's balance of power.

The Deep Game of Technology Roadmaps

The rise of embodied AI is essentially the inevitable result of large model capabilities extending into the physical world. Wang Xingxing noted that for robots to achieve large-scale application, generalization and efficiency issues still need to be resolved, but these will be addressed in the coming years.

At Futian Laboratory, where Liao Hongtao has joined, the research direction is "embodied AI for human living environments," meaning the scenario selection leans more toward home services, elderly care and disability assistance, and other areas close to daily life. If DeepSeek enters the field, how its large model capabilities combine with robot hardware will become a focal point for the industry.

Another noteworthy signal: Skyworth is incubating an exoskeleton product for the elderly, helping seniors alleviate lower limb burden when walking, standing up, and climbing stairs. The project belongs to Skyworth Digital Innovation Research Institute and has gone through multiple iterations but has not yet entered mass production.

From large model companies to home appliance manufacturers, players with different backgrounds are entering the embodied AI track in their own ways.

Generalization is a core technical metric for embodied AI, referring to a robot's ability to stably complete tasks in unseen new scenarios, with new objects, and under new instructions. Traditional robots rely on pre-written rules and fixed programs, and can only work in highly structured environments; embodied AI based on large models attempts to enable robots to acquire general understanding of the physical world through large-scale pre-training, much like large language models process text.

Efficiency involves engineering metrics such as inference speed, energy efficiency ratio, and real-time control. Robots operating in the physical world are extremely sensitive to latency — if an action decision exceeds hundreds of milliseconds, it could lead to grasping failure or safety incidents. Therefore, how to run large models under limited computing power and energy consumption is a threshold that embodied AI must cross to go from demonstration to mass production. Wang Xingxing's remark about "resolving within one to two years" points precisely to engineering breakthroughs in these two categories of problems.

Cool-Headed Thinking: Real Challenges Beneath the Hype

The heat around embodied AI is undeniable, but from technological maturity to a commercial closed loop, many issues remain to be resolved.

The "generalization" Wang Xingxing mentioned refers to whether robots can stably execute tasks across different scenarios, rather than only working under specific conditions. "Efficiency" involves engineering metrics such as inference speed and energy efficiency ratio — none of which can be solved by simply throwing more computing power at them.

The hardware cost of embodied AI also cannot be ignored. Exoskeleton products target the elderly, but whether the price can be accepted by ordinary families will directly determine the market space. If DeepSeek enters robotics, it will likewise need to answer the question of "cost and mass production."

What is certain is that large model capabilities are providing a new technological foundation for embodied AI, and embodied AI has found an outlet for large models to interact with the real world. How far this round of cross-integration can go depends on the speed of technological breakthroughs and the pace of commercialization.

Exoskeleton is a wearable mechanical device that provides assistance to the wearer through sensors, actuators, and control algorithms, commonly used in medical rehabilitation, industrial handling, and military fields. Lower limb exoskeletons targeting the elderly primarily assist with walking, standing up, and climbing stairs, alleviating the burden on knee and hip joints.

Skyworth's incubation of such products reflects the demand gap created by China's accelerating aging population. But the commercialization of exoskeletons faces a dilemma: on one hand, medical-grade products require strict clinical validation and certification, with long cycles and high costs; on the other hand, if consumer-grade products are priced too high, ordinary families cannot afford them, while compressing costs may sacrifice safety and comfort. Skyworth's project "has gone through multiple iterations but has not yet entered mass production" — a microcosm of this predicament.

One-line Summary

DeepSeek founder Liang Wenfeng is in frequent talks with embodied AI founders, planning to enter via investment or in-house development, as the industry's talent war and technology competition intensify.

Paragraph Summary

DeepSeek founder Liang Wenfeng has recently been paying close attention to the embodied AI field, engaging with multiple entrepreneurs, and may assemble a team personally or enter through investment. He previously invested in Unitree Robotics and plans to lead an investment in Winverse AI, founded by former Tencent AI scientist Hu Han, showing an "invest first, then build" strategy. Meanwhile, former UBTech co-CTO Liao Hongtao has joined Tencent Futian Laboratory, and Unitree's Wang Xingxing judges that the embodied AI critical point may arrive within one to two years. The article notes that combining large models with the physical world is becoming a mandatory question for AI giants, but generalization, efficiency, hardware costs, and commercial closed loops remain challenges to be solved.