From WAIC2025 to the World Robot Conference, robots have become the most prominent protagonists. In the AI era, they have also gained a new name - embodied intelligence. This concept has rapidly become one of the most hotly debated tracks in 2025: When will embodied intelligence truly integrate into human life and work? Are humanoid robots the optimal solution for robotics? With these questions in mind, ZHIHU-W (02390) recently launched an online roundtable titled "Advancing Embodied Intelligence," featuring guests including Jiang Zheyuan, Chairman of Songyan Dynamics, Xu Huazhe, Co-founder of Xinghaitu, Zhao Hao, Assistant Professor at Tsinghua AIR Lab, Mo Yilin, Associate Professor at Tsinghua University's Department of Automation, along with numerous frontline practitioners. They engaged in in-depth discussions around cutting-edge technologies, application scenarios, and future trends in the embodied intelligence field, creating what can be called the most comprehensive "embodied intelligence manual" currently available online.
From "Dexterous Faces" to "Uncanny Valley": Industry Professionals Analyze New Directions in Embodied Intelligence
During the discussion, Zhao Hao, Assistant Professor at Tsinghua AIR Lab, shared the team's latest research achievements on "dexterous faces." Zhao explained that the choice to research dexterous faces was based on considerations of differentiated competition and short-term and long-term strategic value: In 30 years, fully-realized humanoid robots with ultra-high emotional value and dexterous faces are expected to become high value-added products in the humanoid robot market.
Zhao introduced that the team's developed Morpheus dexterous face adopts an innovative "hybrid" drive system, where large facial structures are achieved through rigid driven mechanisms, while micro-expressions are accomplished through tendon driven systems. Meanwhile, the team has incorporated the latest digital human technologies, using blendshape-based mapping to enable the dexterous face to produce various rich expressions. This innovative technology application has brought new breakthroughs to robots' anthropomorphic expression capabilities.
However, this advancement brings with it the classic "uncanny valley effect." During this roundtable, the relationship between robots' anthropomorphic degree and the uncanny valley effect also became a hot topic. The uncanny valley theory indicates that as robots' anthropomorphic degree increases, human emotional responses show a curve of initial increase, then decrease, followed by another increase. Zhihu contributor Zero One Monkey believes that one way to mitigate the uncanny valley effect is to avoid making robots too human-like, such as Unitree Technology's approach. However, he also pointed out that in future companion or specific scenarios, robots' faces will still need to approach human likeness.
Another Zhihu contributor, MarkShuo, concisely stated, "Fear doesn't come from robots, it comes from ourselves. Humans are in a constant process of giving up. Only by accepting this can you not be terrified."
From Competition Performance to Academic Theory, Zhihu Builds a Field for Academic-Industry Integration
At the 2025 World Humanoid Robot Sports Conference held in Beijing in August, Tiangong Ultra robot and Unitree H1 won first and second place respectively in the "100-meter robot battle," sparking heated discussions in the Zhihu community. "Beijing Tiangong won with algorithms and autonomous perception, while Unitree Technology excelled in hardware and motion control," commented Zhihu contributor AI Decoder. He believes Beijing Tiangong is more like proving "robots can run independently," while Unitree Technology is validating "robots can run at scale." "If you ask who has more future potential, I'm more optimistic about Unitree's platform-based approach. Because ultimately, the market won't ask how fast you can run, but rather: Can you be cheap, can you mass produce, can you continuously iterate?"
Zhihu contributor Zhufeng praised Unitree H1 robot's obstacle-crossing leap as the most spectacular moment for Unitree robots.
Meanwhile, academic discourse from the academic community has added depth to the "Advancing Embodied Intelligence" discussion. Is control theory outdated? Is it still useful to start learning automation/control theory now? Wang Mengdi, Tenured Professor in the Department of Electrical and Computer Engineering at Princeton University, stated, "Large model alignment and reasoning are essentially control problems, except that the system's inputs and outputs are discrete token sequences, and the control strategy itself equals the model and its parameters."
Tsinghua University Control Engineering Master's degree holder and Zhihu contributor Automation Learning Machine believes, "After AI explores capability boundaries with abundant computing power, it will ultimately discover that control theory is the core, packaged with more gorgeously fashioned neural networks, which is the optimal solution."
Zhihu has built an extremely vibrant technology and AI communication ecosystem in China, gathering 16 million continuous learners in the technology and AI fields, 3.56 million deep content creators in technology and AI topics, and accumulating 8.58 million AI-related questions and 20.88 million professional AI answers. Zhihu not only witnesses the growth and exploration of developers in the AI era but is also becoming a bridge for collaboration between academia and industry, promoting the collision of technical viewpoints and experience sharing, reaching toward a more open and creative AI era.