Why robots dance

China’s robot dancers are not only a new expression of Chinese culture, but also the precursors of embodied intelligence for real-world applications

“Dancing robots” has repeatedly trended on social media after a series of high-profile public performances, including the Yangge dance and Chinese martial arts shows at the Spring Festival Gala. Alongside admiration and applause from the public, there has also been persistent skepticism that this is simply “showing off” or even “useless”.

Why should robots dance?

From a technological perspective, most artificial intelligence today is essentially a “brain in a vat” — skilled at processing symbols, but without a body. Humanoid robots represent the next frontier of AI: embodied intelligence, which allows AI to truly “inhabit” a body and move from the digital world into the physical one.

A robot can be thought of as having its own versions of a human brain, cerebellum and body — the “brain” decides what movement comes next and at what rhythm; the “cerebellum” controls how much force each joint exerts and where the center of gravity shifts; and the “body” consists of motors, sensors and limbs.

For a robot to complete a three-minute dance, its “cerebellum” must make hundreds of real-time decisions every second — continuously shifting its center of gravity, synchronizing more than a dozen joints and staying in time with the music. A delay of just 0.05 seconds in any one of these processes can cause the robot to lose its balance, fail the movement and fall.

An even deeper challenge lies in the transition from simulation to reality. Even a model trained to perfection in the lab can still lose its balance in the real world, because real-world environments are always more complex than any simulation.

Mastering the skill of dance means AI can both “conceive of” a movement and “execute” it. As robots’ ability to learn from training data advances, their movement precision improves accordingly. The more complex the dance movements, the higher the demands on balance and multi-joint coordination — and the stronger the robot’s ability to withstand unexpected disturbances in real environments. A robot that can perform dance movements reliably has the potential to take on more complex tasks, such as industrial inspection, emergency rescue and household services.

In other words, a robot dancing at a heritage site today could be conducting rescue work at a disaster site tomorrow. Dancing isn’t a performance stunt — it’s a rigorous, real-world test.

From the perspective of application scenarios, when a robot operates in a lab, it is answering “questions it has already studied”. When it operates outside the lab, it is stepping into the “final exam” of the real physical world.

In the AI field, there is a specific term: “out-of-distribution”, referring to the real-world situations that robots have never encountered in the training set. Completing a performance smoothly and safely in a real-world setting is the most direct test of an embodied intelligence system’s robustness, or its ability to resist interference. Data collected in real environments is itself valuable training material — it records how a robot responds to unpredictable variables such as lighting, ground conditions and sound, offering samples that are difficult to replicate in a lab and feeding into further algorithm iteration. Through the continual accumulation of this kind of data, robots can gradually progress from completing a single performance to genuinely entering homes, hospitals and communities — serving people’s daily lives and empowering countless industries.

And from a cultural perspective, dancing turns robots into narrators of human civilization.

Choreographing a robot dance of random movements is technically possible, but doing so yields relatively uniform data, and squanders the rich potential of cultural elements to empower new technology.

A team of developers recently completed a translation of cultural elements that bridges a span of about 1,500 years by staging flash mob robot performances at the Yungang Grottoes, a UNESCO World Heritage site. They used AI to convert the dynamic postures of the Northern Wei Dynasty (386-534) guardian statues at the grottoes into data, and then had a fleet of robots physically render that data — so that this piece of cultural heritage could truly come “alive” before people’s eyes.

The process began with dance archaeology, analyzing the physical tension, center of gravity and muscle orientation of 10 sets of the statues at the site one by one. Next came AI-based 3D pose estimation, in which a neural network translated the 2D stone reliefs into 3D skeletal structures. This was followed by motion completion, or in-betweening: since a stone carving captures only a single instant of a movement, AI was used to fill in the preceding and following frames of continuous motion — in keeping with the principles of human biomechanics — so that these static guardian figures could “move”. Next came motion retargeting: since the skeletal proportions of humans and robots differ, each movement had to be precisely mapped before it could be applied to a specific robot body. Finally came physical feasibility verification, in which the movements were tested through simulation before being transferred to the hardware itself.

This work also raises a deeper question worth exploring over the long term: Most AI models today are trained on data from specific cultural backgrounds. Data from English-speaking countries teaches AI to understand and interpret the world through the logic of the English language, leaving such models ill-equipped to grasp the body language of Northern Wei Dynasty guardian statues from China. Using data on Eastern cultural rhythm and posture to train embodied intelligence is a way of building a “cultural body” for Chinese AI — and more than that, it adds Eastern cultural elements to the broader landscape of artificial intelligence, helping align AI with values shared by all of humanity, and promoting exchange and mutual learning among different civilizations.

The significance of embodied intelligence extends beyond any single country or civilization. It represents a technological leap that belongs to all of humanity — one that takes AI out of the virtual “brain in a vat” and into the physical world, while giving the cultural elements of different civilizations an entirely new medium of expression that global audiences can understand.

Film and new media technologies once drove the cross-regional spread of culture and art. Today, AI and robotics go a step further, bringing ancient civilizations “alive” in the real world as tangible, interactive presences. China’s robot dancers are not only a new expression of Chinese culture, but also a new starting point for exchange and mutual learning among civilizations.

Just as Chinese robots can now perform the poses of Yungang’s guardian figures, in the future they may perform Greece’s Hera dance or India’s Bharatanatyam. Such performances would enrich the cultural diversity within the AI world, so that each civilization’s beauty can shine in its own way and, together, create a shared beauty — making robots a new vessel for the diversity of world civilizations and the uniqueness of national cultures.

Tian Qing is a member of the Beijing Dancers Association, a cross-media artist, and the director of the robot dance show at “Macau 2049” and “The Awakening of the Yungang Guardians”. Wang Xingxing is the founder of Unitree Robotics.

The authors contributed this article to China Watch, a think tank powered by China Daily. The views do not necessarily reflect those of China Daily.

Contact the editor at editor@chinawatch.cn.