Galbot and Tsinghua Unveil LATENT: A Humanoid Robot That Learned Tennis from Amateur Clips

Researchers from Galbot (Galaxy General Robotics) and Tsinghua University have unveiled LATENT, a groundbreaking system that teaches humanoid robots to play full-court tennis using only imperfect human motion data. The project, announced on March 15, 2026, represents a major leap in athletic humanoid robotics and reinforcement learning.

Key Highlights

• The humanoid robot learned tennis from just 5 hours of motion capture data consisting of short, fragmented clips of amateur swings
• Achieved a 90.9% success rate on forehand returns and can handle ball speeds exceeding 15 meters per second
• The robot sustains multi-shot rallies against human players and can even rally with another robot
• No onboard cameras or vision models are used; the system relies on external motion capture for high-precision, low-latency ball tracking

How LATENT Works

LATENT, which stands for Learning Athletic humanoid TEnnis skills from imperfect human motioN daTa, solves one of the hardest problems in embodied AI: teaching robots complex, high-speed athletic skills without perfect demonstration data.

Traditional approaches require continuous, flawless 3D tracking data from professional athletes during actual matches, which is extremely expensive and difficult to capture. LATENT bypasses this entirely by using short, disconnected clips of basic human swings as rough movement hints.

The system works in three stages:

1. Motor Cerebellum Construction: The AI uses rough human clips as a foundation and a physics simulator corrects the physical errors, ensuring the robot does not fall over while swinging
2. Motion Skill Space: LATENT creates a unified space of movement skills that the robot can blend and combine to handle various incoming shots
3. Latent Action Barrier (LAB): A safety mechanism that allows the robot to autonomously decide how to return balls while maintaining natural, human-like motion and balance

Why This Matters

The significance extends far beyond tennis. By proving that messy, fragmented human data can be transformed into smooth, highly dynamic robot movements, this research drastically lowers the barrier to teaching robots complex physical tasks.

Previously, training a robot for high-speed physical activities required expensive, purpose-built datasets recorded under controlled laboratory conditions. LATENT shows that readily available, imperfect motion data is sufficient, opening the door for humanoid robots to learn a wide range of athletic and dynamic tasks.

Full-Court Coverage

The robot demonstrates remarkable agility across the entire tennis court. It can cover wide shots, adjust its positioning dynamically, and control where the ball lands on the opponent's side. The whole-body coordination required for each shot, from planting the feet to swinging through the ball, closely mimics natural human movement patterns.

What Comes Next

The LATENT project builds on Galbot's growing reputation in the humanoid robotics space. The company's G1 robot was featured at China's 2026 Spring Festival Gala, and founder Wang He, a Tsinghua and Stanford alumnus, has positioned the company at the intersection of embodied AI and practical robotics.

As humanoid robots continue to demonstrate increasingly athletic capabilities, applications could extend from sports coaching and physical therapy to warehouse logistics and emergency response, anywhere robots need to perform fast, unpredictable physical tasks in real-world environments.

Source: Galbot / Tsinghua University Research