Humanoid Robots Are Learning Tennis — and It’s a Big Leap for Real-World AI

Teaching robots to play tennis might sound like a stunt, but it’s quickly becoming a test of real-world intelligence.

Chinese researchers have developed a new way to teach robots how to play tennis, one that allows them to react in real time rather than follow fixed programming. Using a new system called LATENT, the team trained a robotic model to perform complex, fast, and dynamic tasks with notable accuracy by breaking movements such as strokes and footwork into smaller, learnable components.

Achieving a previously impossible feat

According to the team, earlier attempts have struggled with replicating tennis activities in robots. One reason they mentioned in their abstract for that was the inability to reproduce perfect human kinetic data in tennis scenarios. 

However, by developing LATENT, a model that could work without the perfect kinetic data from human tennis players, they were able to break this barrier. Because these imperfect data still contained core movement skills used in tennis, the model could infer what the overall process should look like by combining those simple elements.

By further correcting the messy data, the team created a control system that allows robots to reliably hit balls under varying conditions while keeping their motions identical to humans’.

In a simpler breakdown, this was what these researchers did:

• They captured and combined bits and pieces from players that contain errors but are easier to capture and break down.
• Used reinforcement learning to help the model figure out, using these data, what the priors were. That means, for instance, the model can understand what a human swing should look like.
• Combined all these into a system called Learns Athletic humanoid TEnnis skills from imperfect human motioN daTa (LATENT), which powers humanoid robots.

Real-world application

Robots have been used in a couple of environments, from commercial use in factories and elsewhere to becoming a physical home assistant. However, robots used in entertainment and sports raise the bar quite high. That is because entertainment requires very nuanced and unpredictable moves.

Tennis is no different. Just as this report from Interesting Engineering puts it: 

“Tennis is especially demanding for humanoids because everything has to happen in one continuous loop. The robot must detect the ball early, predict its trajectory, position its body, stabilize its legs, and swing the racket at precisely the right moment—without stopping to reset. Even a short exchange leaves no room for delay, as the ball keeps moving regardless of processing time.”

Despite all of these, advances have been made in the use of robots for sports. One example is the Walker S2, a multifaceted robot from UBTech that stunned the public with its tennis skills, according to a Jan. 2 report from Interesting Engineering.

In a test conducted with the Unitree G1, the humanoid achieved a peak success rate of 96.5% across 10,000 trials. It made forehand and backhand strokes while sustaining good motion balance, playing against humans.

Improvements and applicability

In the conclusion section of their research paper, the team noted that although their research focused on tennis, further improvements to this system could enable the robot to be used in a variety of entertainment scenarios.

Its current limitations make it a weak match for professional tennis players. But room for improvement exists. Its training setup currently uses randomly generated balls instead of a simulated, true competitive match. 

Also, replacing its motion-capture technology with active vision will improve its real-world operation autonomy. That also includes using a multi-agent training framework that forces each robot to adapt to changes made by another robot during play, thereby simulating real tennis, where opponents can change strategy at any time.

Also read: China’s humanoid robotics sector is expanding quickly, with several new machines showing how far the technology has advanced in 2026.

The post Humanoid Robots Are Learning Tennis — and It’s a Big Leap for Real-World AI appeared first on eWEEK.