Robot that looks like a human is shown playing tennis with people in a video

Can a Robot Compete with Serena Williams?

Researchers have trained a humanoid robot to play tennis alongside humans, and it seems to be doing quite well, actually.

A Chinese AI robotics company named Galbot has developed software that enables the Unitree G1 humanoid robot to engage in tennis play with human engineers.

They shared a video online, showcasing a white robot skillfully holding what appears to be a regular tennis racket, effectively returning the ball while moving across the court.

“Your humanoid tennis player is here!” Galbot proclaimed.

This advancement signifies a shift from merely replicating mechanical movements to more intelligent, decision-making motor interactions.

The software, designated LATENT (Learning Athletic Humanoid Tennis Skills from Imperfect Human Motion Data), is claimed to be the first of its kind capable of real-time whole-body planning and control for athletic tennis.

According to preliminary unpublished papers, the system relies on “incomplete human motion data,” derived from fragmented movements that capture the basic skills necessary for playing tennis, rather than polished data from actual matches.

The short clips of human actions include forehand and backhand swings, plus some fundamental footwork. These snippets create a library of movements that the robot learns to combine in real-time.

In terms of wrist control, the robot’s advanced controller modifies wrist movements directly during play, avoiding reliance on imperfect data.

The robot is capable of engaging in extended tennis matches, responding to balls moving at speeds exceeding 15 meters per second (about 33.5 miles per hour) and executing coordinated strokes with its footwork.

To be honest, the movements look surprisingly natural, considering it’s a robot. While they don’t quite match human fluidity, they don’t come off as stiff or excessively robotic either.

“Our main insight is that, despite being incomplete, this quasi-realistic data still conveys a foundational understanding of human skills applicable in a tennis context,” the researchers noted.

They added, “With further tweaks and setups, we can develop strategies for the humanoid that will consistently strike incoming balls and accurately return them under various conditions, while still appearing natural as it moves.”

In testing, the system achieved an impressive 96% success rate on forehand shots.

However, the engineers also mentioned that this software could extend its usefulness beyond just playing tennis.

“While this study primarily concentrates on the tennis return task, the framework we’ve proposed could potentially apply to a wider array of tasks where comprehensive, high-quality human movement data is lacking,” they suggested.

If robots can master intricate physical tasks like tennis using only imperfect data, it raises the possibility that similar learning methods might be effective for real-world applications too.