Galbot Robotics just rolled out a new humanoid tennis robot. This thing can rally with a human in real time—no scripts, no remote control, just pure back-and-forth.
The robot stands about four feet tall. It’s built on a Unitree G1 platform and runs a system called LATENT.
It watches the ball, tracks it as it zips around, and tweaks its whole-body movement on the fly. The team took a fragment-based approach to teaching it, so it learns movement in pieces—more like how people pick things up in real life.
Real-time tennis rallying: a milestone in sports robotics
Most past demos stuck to pre-programmed routines, but not here. This robot improvises, making split-second decisions as the ball flies by.
The big leap is getting the robot to mix and match small movement fragments, creating new actions on the spot. It’s a lot closer to how humans play than anything before.
Researchers wanted to see if they could turn a little bit of data into flexible behavior. Instead of needing entire matches or endless scripts, they focused on learning from quick snippets of human motion.
How the system works: LATENT on a Unitree G1
The robot’s got a compact, human-ish build. It uses the Unitree G1 chassis and LATENT software to move its whole body together.
For training, they grabbed short movement clips—forehands, backhands, sidesteps—from five different players. All told, that was about five hours of data in a small, 10-by-16-foot court.
They blended these fragments and ran them through simulation, tweaking things like mass, friction, and aerodynamics. The idea was to make sure the robot could handle weird, unpredictable stuff once it hit the real world.
In simulation, it nailed up to 96% success on forehand shots. Out on the court, the robot kept rallies going against a human and managed to return balls across the net most of the time.
Sometimes, it even sent shots away from its partner, which feels like a hint of real decision-making—not just reflexes.
What the trials show in practice
During live play, the robot tracked fast balls and shifted its stance, swing, and timing to return shots. The rallies proved it could react on the fly, not just follow a pre-set routine.
It sometimes misplaces shots, which is actually interesting—it suggests the robot’s starting to make its own choices, not just copy what it’s seen.
The team admits the robot’s motion can look a bit wobbly. Movements aren’t as smooth or efficient as a human pro, and high, unpredictable shots are still tough for it.
Limitations and scope
The results are exciting, but there are some real limits. The robot’s movements aren’t exactly graceful, and it only knows a handful of trained fragments.
High balls, late hits, or sharp angles can still trip it up. It’ll need more data, smarter controls, or better sensors to keep improving.
Broader implications for sport and beyond
This demo feels like a shift toward robots that can really react in fast, interactive settings. The fragment-based learning method seems like it could work for all kinds of physical systems that need to stay flexible.
Expansion to other sports and non-sport tasks
People in the field think this approach could show up in other sports—or even outside sports where quick, context-aware moves matter. Some possibilities:
- Football (soccer): robots making real-time passes and controlling the ball
- Badminton and table tennis: where timing is everything
- Industrial work: adapting to shifting weights and surfaces on the fly
- Search and rescue: robots reacting to unpredictable scenes
Looking ahead: exhibitions, training, and safety
The team’s thinking about live exhibition matches, using the tech for athlete training, and pushing fragment-based learning further in research. Safety, reliability, and making sure we understand what the robot’s doing will be big as these bots move out of the lab and into the real world.
Takeaways for researchers and industry
This work marks a real step forward in dynamic reacting robotics. It shows that robots can actually learn to operate from partial human motion data and turn that into real-time, goal-driven actions.
The LATENT approach gives us a framework for building more adaptable agents. Think sports, industry, even rescue missions—there’s a lot of potential here.
Here is the source article for this story: Robot plays tennis with humans in real time