Humanoid Learns Tennis from Imperfect Data: AI Achieves Realistic Gameplay with Unitree G1

BEIJING – Researchers at Tsinghua University in China have unveiled “LATENT,” a system designed to teach humanoid robots athletic skills, specifically focusing on tennis. The project, announced Tuesday, tackles a significant challenge in robotics: replicating complex human movements with limited, imperfect data. LATENT utilizes a learning pipeline encompassing pre-training, knowledge distillation and high-level policy learning, all powered by the MuJoCo physics engine.

The core innovation lies in LATENT’s ability to learn from incomplete motion data. Traditionally, training a robot to perform a skill like a tennis swing requires extensive, precise recordings of human athletes. However, acquiring such datasets is difficult, and expensive. The Tsinghua team circumvented this issue by using “motion fragments” – basic skill components – rather than complete tennis sequences. This approach dramatically reduces the complexity of data collection.

“Humans utilize versatile and highly dynamic tennis skills to return high-speed tennis balls,” the researchers explained in their project documentation. “However, reproducing such movements in a humanoid is extremely difficult due to the lack of perfect humanoid motion data or human kinematic motion data as references for tennis.”

LATENT’s approach isn’t about flawless replication from the start. Instead, it leverages the incomplete data as a foundation, then employs correction and synthesis techniques to create a natural-looking and effective playing style. The system aims to train humanoids capable of consistently returning balls under a variety of conditions and directing them towards specific targets.

The project’s success extends beyond simulation. Researchers have developed designs to facilitate robust data transfer from the simulated environment to the real world, and have successfully demonstrated LATENT’s capabilities on a Unitree G1 robot. A video released alongside the announcement shows the robot playing a rally with a human opponent, a significant step toward more sophisticated robotic athleticism. You can view the demonstration here.

The underlying technology relies heavily on MuJoCo, a free and open-source physics engine designed for research and development in robotics and biomechanics. MuJoCo offers a unique combination of speed, accuracy, and modeling power, making it ideal for complex simulations like those required for training a robot to play tennis. The system also incorporates techniques like “knowledge distillation,” where information from a more complex model is transferred to a simpler one, and “pre-training,” which allows the robot to learn basic skills before tackling the full task.

The LATENT project is currently open-source, with the tracking codebase and a subset of human tennis motion data released on March 13, 2026, according to the project’s GitHub repository. Further releases are planned, including the full human motion dataset, pre-trained trackers, and the codebase for online distillation and high-level policy learning. This open-source approach is intended to foster collaboration and accelerate progress in the field of robotic learning.

This development arrives at a time of increasing interest in the intersection of artificial intelligence and sports. From AI-powered analytics to robotic training partners, the potential applications are vast. While a robot challenging a professional tennis player remains firmly in the future, LATENT represents a tangible step toward that possibility. The ability to learn from imperfect data is particularly crucial, as it mirrors the way humans often learn – through trial and error, and by building upon incomplete information.

For those following the advancements in AI and robotics, the LATENT project offers a fascinating glimpse into the future of athletic training and robotic capabilities. The team’s success in teaching a robot to play tennis, even with limited data, demonstrates the power of innovative algorithms and sophisticated simulation tools. The open-source nature of the project also promises to accelerate further research and development in this exciting field.

The next step for the Tsinghua team is the planned release of the complete human tennis motion data and pre-trained trackers, which will allow other researchers to build upon their work. The release of the high-level policy learning codebase will also be a key milestone, enabling the development of more sophisticated robotic tennis players. Keep an eye on the LATENT project website for updates and further information.

What are your thoughts on robots learning to play sports? Share your comments below!