Silky-Smooth Play and a Packed Venue
By Tingyu, reporting from Aofeisi Temple
Robots have started playing table tennis too!
This time, the table was brought straight to the World Robot Conference (WRC).
At the Chaowei Power booth, a full-size humanoid robot watched an incoming table-tennis ball, swung its paddle, rotated its waist, and rapidly adjusted its movements based on where the ball landed.
After just a few rounds, a crowd quickly gathered around the booth.
Next to the table, an even larger system was also put on display:
The full-size humanoid KAI Bot, the high-degree-of-freedom dexterous hand KAI Hand, the data-collection headband KAI Halo, KAI World Model, and KAI Embodied AI Infra, which covers the entire data and training workflow.
From hardware and models to data collection and infrastructure, this complete full-stack embodied-intelligence system was brought to the venue.
The setup is bound to remind people of Figure. As one of the world’s highest-valued embodied-intelligence companies, Figure is almost certainly the most closely watched star of this humanoid-robot boom.
What won over investors was not simply the fact that it built a humanoid robot. It was the company’s simultaneous bet on the robot itself, the model, and real-world scenarios.
Figure builds its own robots, trains its own Helix model, and then sends those robots into BMW factories. Data generated by real-world tasks flows back into the model, creating a full-stack loop that runs from the robot body and model to real-world deployment and data feedback.
Chaowei Power is taking the same full-stack approach.
This Chinese company, founded roughly a year ago, has already put its complete robots, dexterous hands, data-collection equipment, world model, and training infrastructure on the table at the same time.
But why must these pieces exist together, and how can they truly interlock into a continuously evolving system?
The answer lies in the full-stack path Chaowei Power has chosen.
In Embodied AI, the Full-Stack Approach Is Becoming the Consensus
Over the past few years, the easiest way for a humanoid robot to go viral was to perform a sufficiently flashy movement.
Running, jumping, or folding clothes—a video lasting just a few dozen seconds could attract huge amounts of traffic.
But people are no longer looking only at demos.
A robot completing a task once merely proves that the task can be done. Whether it can reproduce that performance reliably across different environments and on different robot bodies is what determines whether it has a chance of becoming a product.
Around the world, leading players may have started from different places, but they are all filling out their full-stack capabilities.
Figure entered through the “robot body + large model” combination. It is developing the Helix VLA model in-house while sending robots into BMW factories to accumulate experience through real-world tasks and feed the resulting data back into the model.
Tesla Optimus’s trump card is Tesla’s vertically integrated capabilities.
The algorithms accumulated through FSD, a mature supply chain, and deployment scenarios provided by Tesla’s own factories allow it to advance robot development, model training, and mass manufacturing simultaneously.
Physical Intelligence (π) started from the model side, focusing on general-purpose embodied foundation models. It is attempting to make a single model work with different robots, providing general-purpose manipulation capabilities for robot bodies of varying forms.
The approaches differ, but the goal is becoming increasingly consistent: connect the robot body, model, data, and real-world scenarios in a single iteration loop.
Chaowei Power’s choice is to deploy across hardware platforms, embodied foundation models, high-quality data collection, and underlying infrastructure at the same time.
Luo Ping, co-founder of Chaowei Power and a professor at the School of Computing and Data Science at the University of Hong Kong, offers a straightforward explanation:
The users of language models are ordinary people, while the users of the “brain” and “cerebellum” of embodied systems are robots.
The two types of models receive entirely different kinds of feedback. A language model outputs tokens, whereas an embodied model outputs the robot’s action in the next second. An error does not merely mean an inaccurate answer; it may result in a failed grasp, a collision, or a loss of balance.
This also determines that embodied intelligence cannot learn solely from internet data. Models need real-world action data, and they also need to be repeatedly trained and validated on robot bodies.
Data enters the model, the model drives the robot body, and the results of the robot’s actions return to the training system. If any link is missing, the loop is broken.
From this perspective, going full-stack is not about a company trying to broaden its business scope. It is an inevitable requirement once embodied intelligence enters the real world.
How Do Hardware, Models, and Data Interlock to Form a Closed Loop?
Chaowei Power’s full-stack capabilities ultimately need to materialize in concrete products.
The problem it truly wants to solve is not simply how to build a robot, but how to turn human experience into robot capabilities more smoothly.
First, Build a Body That Can Carry Human Data
Chaowei Power’s full-size humanoid robot, KAI Bot, stands 173 centimeters tall, weighs 70 kilograms, and has 117 degrees of freedom across its body.
Nearly 80% of its body is covered by tactile skin, integrating approximately 18,000 tactile sensing points. In theory, it can detect light touches of 0.1N or more.
The significance of these specifications is not merely that they make the robot’s movements look more human.
More importantly, robot training increasingly depends on human demonstration data. But if a person needs to use their shoulders, waist, and hands to perform an action that a robot’s structure cannot physically reproduce, no amount of data can be mapped directly onto the robot body.
During motion retargeting, every missing degree of freedom forces the model to make another compromise.
Luo Ping says that the closer a robot’s physical structure is to the human body, the more easily it can learn how people operate from first-person data.
Tables, chairs, tools, and spaces in human society are likewise designed around the human body. A high degree of anthropomorphism is not about appearance; it is about lowering the barrier to reusing human experience.
This is the core logic behind KAI Bot’s pursuit of a high degree of freedom: degrees of freedom are not just parameters to stack up. They effectively determine whether the model can fully execute the movements it has learned.
The hands handle the most delicate physical interactions. Sold separately, KAI Hand has 37 degrees of freedom, including 20 active degrees of freedom, one passive degree of freedom, and 16 compliant degrees of freedom. Its fingertip force exceeds 30N.
After continuous gripping for 10 minutes, the maximum temperature of the entire hand remains below human body temperature.
High degrees of freedom provide the space for movement, compliant structures absorb collisions, and strength and thermal management determine whether the hand can work for extended periods.
Together, these four elements form a hand truly designed for embodied models.
From SMASH to KAI World Model: How Robots Understand and Predict the Real World
If KAI Bot and KAI Hand answer the question of whether the body is capable enough, KAI World Model addresses how a robot understands the physical world and predicts the consequences of its own actions.
Luo Ping compares it to a real-time-generated VR world: people use both hands to manipulate objects in a generated environment, while the model generates subsequent states based on their actions and first-person visual input. The resulting data is then used to train humanoid robots.
This approach does not rely on a single technology. The team is simultaneously exploring JEPA-style methods, 3D approaches, and video generation, collectively describing the effort as 4D world modeling.
But generating an interactive environment is only the starting point. A world model must ultimately return to the real world, where it is tested against speed, collisions, and differences between robot bodies.
The SMASH humanoid-robot table-tennis system is a vivid demonstration of KAI World Model’s capabilities.
Table tennis involves high-speed balls, constantly changing spin, and random landing points. A robot cannot complete the task using preset movements alone. It must continuously observe its environment, predict the incoming ball, and adjust its next movement in real time based on that prediction.
Specifically, SMASH does not rely on a single large model to handle everything from scratch. Instead, it connects visual perception, trajectory prediction, motion planning, and whole-body control into a closed-loop system.
The system first captures the incoming ball and predicts its trajectory within milliseconds. It then continuously fuses visual data with high-frequency body-state information to calculate the timing, speed, and angle of each shot.
What is even more notable is that SMASH recently announced its latest milestone: “The first complete autonomous humanoid-robot table-tennis match in human history!”
The robots played out a complete 11-point match against each other.
The next step predicted by the system must be translated into action.
The team collected dozens of hours of human motion data, enabling the robot to learn how to hit the ball from human movements. It then uses a whole-body coordination strategy to engage the shoulders, elbows, wrists, waist, and legs in completing the shot.
More importantly, SMASH has not been deployed only on the company’s own robot bodies. Its algorithms have also been adapted to robots including Unitree G1 and AGIBOT Expedition A3.
From seeing and predicting to acting, the table-tennis demo showcases the closed-loop capabilities required by a world model—as well as the possibility of reusing those capabilities across different robot bodies.
The Data Loop Determines How Quickly a Model Can Evolve
Once the body is ready, the next step is to provide the model with a continuous stream of real-world experience.
Embodied-intelligence data cannot simply be scraped like text from webpages. How people open cabinet doors, pick up items in supermarkets, or adjust the force applied when their fingers touch an object all need to be recorded again in real-world settings.
KAI Halo is Chaowei Power’s gateway to data from the real world.
Hundreds of data collectors wear the equipment and enter homes, supermarkets, commercial spaces, and small factories to gather first-person multimodal data.
The company says it has accumulated more than 100,000 hours of video data, covering over 20 scenarios and more than 300 full-body atomic skills.
The data is further processed to generate full-body human joint trajectories, three-dimensional scenes, hand poses, and action-semantic labels, minimizing the need for repetitive processing before the data enters the training pipeline.
But collecting the data is only the beginning. As millions of clips continuously enter the backend, iteration will quickly grind to a halt if processing, training, and evaluation still depend on manually stitching everything together.
KAI Embodied AI Infra handles the second half of the process. It covers data processing, model training, simulation-based evaluation, and deployment on real robots, while also sending the robots’ operating results back into the system.
According to official figures, the platform can increase the efficiency of high-quality data production tenfold and boost model training and evaluation speeds by an average of five to seven times.
At this point, KAI Halo handles collection, Infra handles processing, World Model handles learning, and KAI Bot handles initial validation and algorithm deployment. Different robot bodies then carry out the tasks.
Feedback from the real world enters the next round of training, and only then do the five pieces truly interlock.
When funding and computing power are not overwhelming advantages, closed-loop efficiency becomes an important lever for Chinese startups to narrow the gap.
One Year to Build a Technical Framework Designed to Mature in Three
Looking across the Pacific, Figure was founded in 2022. It spent several years gradually filling out its robot body, Helix model, and factory deployments before forming the relatively complete loop it has today.
Chaowei Power has moved much faster.
The company was founded in Shenzhen in July 2025. In roughly one year, its complete robot, dexterous hand, data-collection equipment, world model, and training infrastructure have all been brought to fruition. The core pieces required for embodied intelligence have already taken basic shape.
This pace is also related to the team’s accumulated experience.
Chaowei Power’s team spans embodied models, computer vision, motion control, autonomous driving, and robotic hardware. Its members have previously worked on projects including the mass production of medical rehabilitation exoskeletons, the deployment of Level 4 autonomous mining trucks, and end-to-end autonomous-driving models.
The company is also collaborating with research teams at HKU MMLab, Shenzhen Hetao Institute, and other institutions to advance research into embodied models and motion control.
However, expanding the product line is only the first step.
Humanoid robots will ultimately have to prove their reliability over long periods of operation, their ability to be delivered at scale, and whether they can create value after entering real-world scenarios. For a company founded only about a year ago, advancing the robot body, model, data, and infrastructure simultaneously will place considerable pressure on both its finances and organization.
Chaowei Power therefore did not wait for the entire system to mature before considering commercialization. Instead, it first separated out the relatively independent capabilities within the system.
During WAIC 2026, KAI Hand and KAI Halo Lite began public sales.
Beyond the complete robot, the dexterous hand can be integrated into other robots, the data-collection equipment can serve model teams, and Infra can also be offered as a technology service.
Each standalone product is a commercial entry point while also bringing more data and scenarios back into the company’s internal loop.
This is what makes Chaowei Power particularly worth watching.
Although it trails Figure in funding, accumulated real-world scenarios, and commercialization progress, the two companies have highly similar views of the industry’s ultimate direction:
The competition in embodied intelligence will not be merely about building a particular robot. It will be about who can get the robot body, model, data, and real-world scenarios moving faster together.
Chaowei Power’s advantage lies in the fact that its full-stack strategy was established all at once and remains highly focused. In roughly one year, it has not only built the framework for an entire technology stack but also found market entry points for several of its capabilities.
If these products can bring data back to one another,