Robot rallies tennis balls with humans and actually holds its own
A robot built on the Unitree G1 platform is now capable of sustaining multi-shot tennis rallies with humans, reacting to balls travelling over 15 metres per second and returning them to target areas. The system, called LATENT, learns from imperfect human motion data rather than clean motion capture, and still manages to produce coordinated strokes and footwork.
How the robot actually learned to play
The interesting part is not just that the robot can hit a tennis ball. It is how it learned to do it.
Instead of relying on perfect motion capture data, the researchers used short fragments of human movements. Things like forehand swings, backhand strokes and basic footwork. These fragments were not precise, and they did not represent full tennis rallies.
The system stitches together these fragments using a learned latent action space. In simple terms, it builds a library of movement building blocks and then figures out how to combine them in real time. So instead of copying a textbook forehand, it learns something closer to how a human moves when playing tennis, then refines that into something usable.
There is also a clever workaround for one of the hardest problems in tennis, wrist control. The robot’s high-level controller directly adjusts the wrist during play, rather than relying on the imperfect training data.
That alone tells you how messy this problem is. Even humans struggle to explain what their wrist is doing mid-swing.
What the video actually shows
The video doing the rounds online makes this look almost match-ready, but it is worth looking a bit closer.
The robot is genuinely tracking the ball and returning shots. This is not stitched footage or selective cuts. But the rally is clearly controlled. The human is feeding relatively clean, predictable balls rather than pushing the pace or mixing things up.
But it’s still impressive. And fairly surreal.
The system works because it can anticipate fairly consistent trajectories. You do not see last-second adjustments or recovery from awkward bounces. The swings are functional, not precise, and there is a slight hesitation between movements that would become more obvious if the tempo increased.
What does stand out is the coordination.
The footwork is surprisingly decent. The robot repositions itself, shuffles across the court, and times its swings in a way that feels closer to natural movement than earlier humanoid demos. It is not fluid in a human sense, but it is not rigid either.
What it can actually do on court
This is not just a one-hit demo.
The system can sustain multi-shot rallies with human players and return balls across different areas of the court. The incoming balls in testing travel at speeds above 15 metres per second. Which is more basic level tennis. But if this is the worst its ever going to be, it makes you wonder.
In simulation, the robot handles thousands of trials and consistently returns balls close to target areas. In real-world testing, performance holds up across forehand and backhand strokes, as well as different court positions.
There is also another interesting detail. When two of these systems play each other in simulation, they can keep rallies going for up to 25 consecutive shots.
Why this is more interesting than it looks
Robots playing sports is not new. There have been table tennis bots, badminton experiments, even robot football teams. What stands out here is the approach.
Most systems depend on clean, high-quality motion data or unrealistic physics assumptions. This one leans into imperfect data and still produces movement that looks relatively natural.
That has broader implications. If a robot can learn a complex physical skill like tennis from messy, incomplete data, it suggests the same approach could work for other real-world tasks. Anything involving coordination, timing, and adaptation.
Also, instead of trying to perfectly model physics, the system trains with a wide range of variations. Friction, mass, air drag, all of it gets randomised. That makes the robot more robust when moving from simulation to reality.
Where this could go next
Before anyone starts worrying about being replaced at Wimbledon, there are limits.
The setup still relies on motion capture systems to track the robot and the ball. The rallies are controlled, and the robot is not diving for impossible shots or dealing with real match pressure.
The next step is moving away from motion capture and toward vision-based systems, so the robot can operate independently. Another direction is training in more realistic match conditions, with unpredictable shots and longer rallies.
Subscribe to our monthly newsletter! Check out our YouTube channel.
And of course, you can follow Gadgets & Wearables on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds.
