World's First Humanoid Robot Autonomous Table Tennis Match Debuts at 2026 World Robot Conference

The SMASH 2.0 system was demonstrated in a standard table tennis interactive area at the booth, where visitors could play against the robot. The system integrates high-speed visual perception, real-time trajectory prediction, whole-body motion control, and embodied AI decision-making. It can recognize incoming balls, calculate trajectories, and plan swing motions within millisecond-level windows, performing autonomous serves and returns while hitting long balls and drop shots. Leiphone reported that the core algorithm is cross-embodiment deployable, meaning the same algorithm can adapt to different robots.

The KAI world model, described as the robot's "brain," follows a three-step closed loop: generate a virtual world for action interaction, evaluate through models, and then learn the rules of the physical world. It supports both generation and reconstruction tasks, enabling cross-task and cross-embodiment generalization. The model has been pre-trained on millions of videos, which significantly improves its performance after transfer to real robots. The company also demonstrated fine manipulation tasks such as autonomous bottle opening with robotic arms.

The KAIBot, featuring 117 whole-body degrees of freedom, walked around the exhibition area and interacted with visitors. It used a built-in large model to answer questions in real time, performed precise object grasping, and delivered gifts. Superdynamic KAI stated that a sufficiently humanoid body is necessary to quickly integrate into human life and work scenarios and to better inherit human motion data.

The KAI Hand is a bionic dexterous hand with 37 degrees of freedom (20 active, 1 passive, and 16 compliant), sized 1:1 to a human hand. Its passive compliant design provides automatic buffering on impact, and it can sustain fingertip forces above 30 Newtons for continuous grasping. It is equipped with five fingertip tactile sensors and a wrist visual sensor, compatible with mainstream embodied open-source datasets. At the booth, visitors could compare palm sizes with the hand and hit it with wooden boards to test its robustness.

The KAI Halo is a first-person data collection headband with four cameras that fuse multi-camera vision and IMU data to achieve centimeter-level whole-body pose understanding and environment reconstruction. Weighing about 300 grams, it supports long-duration recording and provides a complete data collection platform covering task assignment, cloud processing, model training, and deployment.

The technology foundation is the KAI Embodied AI Infra platform, which unifies the entire pipeline from data acquisition to real-robot deployment. According to Leiphone, it increases data production efficiency by 10 times, accelerates training by 40%, and speeds up model evaluation by 7 times. With large-scale cluster scheduling and 5x data loading speed, training speed improves by an average of 5 times. The platform supports unified management of millions of devices and can push models to robots within milliseconds.

The conference runs from August 19 to August 23 at the Beijing Yichuang International Convention and Exhibition Center, where Superdynamic KAI is located at booth B215. The company said it aims to present a complete set of full-stack embodied intelligence capabilities, from the "brain" to the "body."