Chinese Startup Super Dynamic Shows Full-Stack Embodied AI with Table-Tennis Robot at WRC

At the booth, KAI Bot tracked incoming balls, swung its paddle, rotated its waist and adjusted its posture in real time. Beside it, the company showcased KAI Hand, a high-degree-of-freedom dexterous hand; KAI Halo, a data-collection headset; KAI World Model; and KAI Embodied AI Infra, a platform covering data processing, model training, simulation evaluation and real-robot deployment.

Super Dynamic’s approach mirrors that of Figure, one of the most valuable embodied AI companies. Figure develops its own humanoid hardware, trains the Helix model, and deploys robots in BMW factories so that real-task data flows back to the model. Tesla Optimus relies on vertical integration, while Physical Intelligence focuses on a general-purpose embodied foundation model. The strategies differ, but the industry consensus is converging: linking the body, model, data and real-world scenarios in one iterative loop.

Luo Ping, co-founder of Super Dynamic and professor at the University of Hong Kong’s School of Computing and Data Science, explained why a full stack is necessary. "The users of language models are ordinary people, while the users of embodied models are robots," he said. An embodied model outputs a robot’s next action rather than a token, so an error can cause a failed grasp, a collision, or a loss of balance. Embodied intelligence cannot learn only from internet data; it needs real-world action data and repeated validation on physical robots.

The company designed KAI Bot with a height of 173 cm, a weight of 70 kg, and 117 degrees of freedom. Nearly 80 percent of its body is covered with tactile skin containing about 18,000 tactile points, theoretically capable of sensing touches of 0.1 N or lighter. Luo stated that the closer the robot body is to human structure, the easier it is to learn manipulation from first-person human data. The high degree of freedom is not for parameter stacking; it determines whether a learned action can be fully executed.

KAI Hand, sold separately, has 37 degrees of freedom: 20 active, one passive and 16 compliant. Its fingertip force exceeds 30 N, and after 10 minutes of continuous gripping, its maximum temperature stays below human body temperature. The high DOF provides action space, the compliant structure buffers collisions, and power and thermal management decide whether the hand can work for long periods.

KAI World Model aims to help robots understand the physical world and predict the consequences of their actions. Luo compared it to a real-time generated VR world where a human manipulates objects, and the model generates subsequent states. The team explores JEPA-style approaches, 3D methods and video generation, collectively described as 4D world modeling. The SMASH table-tennis system is a direct demonstration of this capability. It combines visual perception, trajectory prediction, action planning and whole-body control, capturing the ball in milliseconds and computing stroke timing, speed and angle. Super Dynamic said SMASH recently completed what it called the first full 11-point table-tennis match between two autonomous humanoid robots. The system has also been adapted to Unitree G1 and Agibot Yuanzheng A3 robots.

Data is collected through KAI Halo, worn by hundreds of personnel in homes, supermarkets, commercial spaces and small factories. The company reported over 100,000 hours of first-person video data, covering more than 20 scenarios and 300 whole-body atomic skills. The data is processed into joint trajectories, 3D scenes, hand poses and semantic labels. KAI Embodied AI Infra then handles processing, training, simulation and deployment, claiming a 10-fold improvement in high-quality data production efficiency and a 5- to 7-fold average speedup in training and evaluation. The loop closes when robot execution results feed back into the system.

Super Dynamic was founded in Shenzhen in July 2025. In about one year, it has built the core pieces of a full-stack system. The team spans embodied models, computer vision, motion control, autonomous driving and robot hardware, and its members have worked on mass-produced medical rehabilitation exoskeletons, L4 mining trucks and end-to-end autonomous driving models. According to an unofficial tip, the Infra team previously built the infrastructure behind Huawei’s autonomous driving from zero to one million mass-produced vehicles. The company also collaborates with the University of Hong Kong’s MMLab and Shenzhen Hetao College.

Commercialization has also begun. At WAIC 2026, KAI Hand and KAI Halo Lite were put on public sale. The company believes that each independent product can serve as an entry point while bringing back more data and scenarios for the internal loop. Compared with Figure, which was founded in 2022, Super Dynamic has moved faster in assembling its stack, though it still trails in funding, scenario accumulation and commercialization. The key test ahead is how quickly the closed loop can accelerate, which will determine whether Chinese embodied AI startups can catch up with or surpass Silicon Valley giants.