Chinese researchers trained a robot to play tennis in an innovative way

Chinese scientists unveil a new method for training robots to play tennis

Researchers from China have published results of testing an innovative methodology that enables robots to quickly and simply master the basic skills of playing tennis. According to their assessment, this could represent a significant breakthrough both in machine learning and real-world AI applications – reports New Atlas.

Why traditional technologies don’t work
In most sports, including tennis, motion capture systems are still unable to record the finest details, such as wrist angle at impact. On a dynamic court such nuances are critical, and remote control proves ineffective.

The problem is compounded by attempts to extract the necessary information from multi‑camera video recordings using AI software (e.g., Nvidia’s Vid2Player3D). This “complex process” requires deep knowledge and engineering effort.

What the researchers proposed
They created a system called LATENT based on motion capture, but limited only to basic elements of technique. Such a system can operate with incomplete data.

- Experiment: over five hours, data were collected on “primitive skills” – right/left swings, lateral movements, and cross‑steps on a partial court.

- This data was processed by cameras to create a repertoire of human‑like “movement spaces”.

- Then the basic skills were loaded into the humanoid robot G1 from Unitree (cost $13,500).

How the robot learns
The LATENT system allows G1 to recognize an incoming ball and, using a racket, return it over the net. Success is counted when the ball lands within the white lines on the opponent’s side of the court.

The robot uses basic skills to experiment with angles, reaction time, and movement choices in different situations. Most of the training occurs in high‑speed simulation.

Results
- 90 % success rate for right swings.

- ≈80 % for left swings.

- Movements appear smooth and agile, almost like a real tennis player.

Although G1 is not yet ready for official matches, it has already demonstrated significant progress in mastering the game.

What this means for the future of robots
The developed method allows robots to quickly adapt to complex and dynamic situations. This opens prospects for practical tasks that require rapid response to extreme conditions – from industrial production to rescue operations.

LATENT software is open source and available on GitHub.