Unitree G1 Humanoid Robot Achieves Advanced Tennis Proficiency Using Minimal Data Sets

By the year 2026, the field of robotics has witnessed a monumental breakthrough in teaching humanoid systems to perform complex dynamic tasks. Researchers from Tsinghua University and Peking University, in collaboration with the technology firm Galbot, have unveiled the LATENT project. This initiative enabled the Unitree G1 humanoid robot to acquire sophisticated tennis skills by training on just five hours of incomplete motion capture data sourced from amateur players, marking a significant departure from traditional data-heavy training methods.

During performance evaluations, the Unitree G1 demonstrated exceptional precision, reaching a successful ball return rate of approximately 90.9% when performing forehand shots. These trials involved human opponents who served balls at speeds exceeding 54 km/h. The foundation of this achievement is the "Correctable Latent Action Space," a framework that provides stabilization to the robot's movements during high-speed swings and impacts. The significance of this milestone lies in its ability to bypass the traditional requirement for vast quantities of high-fidelity data, potentially accelerating the deployment of humanoids in real-world environments.

The development team successfully optimized the training process by reducing the required motion capture area by more than 17 times compared to the dimensions of a standard tennis court. The research focused on fundamental athletic skills, including forehand and backhand strikes, alongside essential footwork like side-steps and cross-steps. In simulation environments, the G1 achieved a 96% success rate for forehand shots, while physical trials confirmed its capability to return balls traveling at peak velocities surpassing 15 meters per second.

The LATENT framework is unique in its ability to utilize "imperfect" motion fragments as a structural prior to refine agile and natural movements. However, technical disclosures also highlighted current limitations of the system. At this stage, the Unitree G1 operating under the LATENT framework primarily returns balls reactively and does not yet employ strategic shot placement against human competitors. Furthermore, ball-tracking during these tests relied on an external optical motion capture system rather than the robot’s integrated vision sensors.

The researchers intend to expand the application of this framework to other high-intensity activities such as soccer and parkour, although autonomous strategic decision-making remains a distinct challenge for future development. Within the broader context of 2026, this methodology offers a vital scalable path for robotic training. Unitree, the Hangzhou-based manufacturer, has announced plans to ship as many as 20,000 humanoid robots this year, representing a sharp increase from the 5,500 units delivered in 2025.

The Unitree G1 robot used in this project is available in various configurations offering between 23 and 43 degrees of freedom (DoF), with prices ranging from $13,500 to $43,000. This advancement is part of a larger trend in 2026 where Unitree has demonstrated its robots performing diverse skills such as dancing and martial arts. This shift underscores the industry's drive to move humanoid technology out of specialized laboratories and into practical, utilitarian roles. Galbot is already spearheading this transition in the retail sector, currently managing over 40 autonomous stores throughout China using similar advanced robotic capabilities.