Physical AI
Galbot's Robot Plays Tennis With Humans: Inside the LATENT System That Learned From 5 Hours of Amateur Data
Galbot Robotics demonstrated a Unitree G1 humanoid robot sustaining real-time tennis rallies with a human player using the LATENT framework. Trained on just five hours of amateur motion data captured in a 3x5 meter room, the system achieved 96% forehand accuracy in simulation and showed natural whole-body coordination on a real court.
A humanoid robot just played tennis with a human, and it held its own. Galbot Robotics, a Beijing-based startup backed by nearly $900 million in funding, released a video on March 16 showing a Unitree G1 robot sustaining real-time rallies against a human player. The robot returned shots traveling at over 15 meters per second (about 34 mph), using an unmodified tennis racket and whole-body coordination that looked surprisingly natural.
The system behind it is called LATENT (Learning Athletic Humanoid Tennis Skills from Imperfect Human Motion Data), developed in collaboration with researchers from Tsinghua University and Peking University. In simulation testing, it achieved a 96% success rate on forehand returns. The real significance isn't tennis. It's that a robot learned a fast, dynamic physical skill from just five hours of imperfect human motion data, then transferred it to a real body.
AI-generated image
The LATENT system was tested on a real tennis court with human opponents. Source: Galbot Robotics
Key Stats
96%
Forehand Success (Sim)
15 m/s
Ball Speed Handled
5 hrs
Training Data Used
$887M+
Galbot Total Funding
How LATENT Teaches a Robot to Play Tennis
Tennis is one of the most demanding athletic tasks you can throw at a robot. The ball can travel at speeds up to 30 m/s, racket-ball contact lasts just a few milliseconds, and the player needs to coordinate legs, torso, arms, and wrist in real time while reading the ball's trajectory, spin, and bounce. Traditional robotics approaches would require exhaustive motion capture data from professional players in controlled environments. Galbot's team took a different path.
The LATENT framework starts with something counterintuitive: imperfect data. Instead of recording full tennis matches with expensive motion capture rigs on a standard court, the researchers captured short movement fragments from five amateur players in a space just 3 by 5 meters, roughly 17 times smaller than a regulation court. They recorded forehand swings, backhand strokes, lateral shuffles, and recovery steps. Total recording time: about five hours.
These fragments become a library of "primitive skills." The system's first training phase teaches the robot to reproduce each movement individually. The second phase stitches these primitives together into sequences: track the ball, move into position, execute the stroke, recover to a ready stance. A high-level controller manages real-time decisions about which primitive to deploy and when, while a separate wrist controller handles the fine racket-angle adjustments needed for accurate ball placement.
🎯 Primitive Skill Library
Short motion fragments (forehands, backhands, footwork) captured from five players in a 3x5m space. These form the building blocks the robot combines during play.
🔄 Domain Randomization
The model trains in simulation with randomized physics parameters (mass, friction, aerodynamics) to close the gap between virtual training and real-world deployment.
Under the Hood: From Simulation to Real Courts
The technical challenge LATENT solves is what robotics researchers call the sim-to-real transfer problem. A model that works perfectly in simulation often fails on a physical robot because the simulation doesn't perfectly capture real-world physics. LATENT addresses this through aggressive domain randomization: during training, the simulation randomly varies the robot's mass, joint friction, ball aerodynamics, court surface properties, and sensor noise. The robot learns policies that are robust across a wide range of physical conditions, not just one idealized setup.
The system architecture splits into two layers. A high-level policy running at lower frequency handles strategic decisions: where to move, when to begin the swing, which stroke type to use. A low-level controller running at high frequency handles joint-level torque commands and real-time balance. The wrist controller operates semi-independently, making micro-adjustments to racket angle that the imperfect training data couldn't capture. This separation lets the robot combine learned human-like motion with precise mechanical control.
AI-generated image
Motion tracking systems capture the primitive skill data that forms LATENT's training foundation.
The paper, posted to arXiv on March 16 and not yet peer-reviewed, reports that in simulation the system achieved up to 96% forehand success rates and maintained consistent ball placement to target zones. On the physical Unitree G1 robot, the system demonstrated sustained multi-shot rallies with a human opponent. The robot reacted to balls traveling at 15+ m/s, executed coordinated strokes, and moved laterally across the court to reach wide shots.
Key Insight
The researchers' core finding is that "imperfect" motion data, the kind you can collect cheaply in a small room, still provides useful priors about human movement patterns. You don't need perfect data. You need enough signal for the robot to learn the right primitives, then let simulation and domain randomization handle the rest.
Platform Comparison: Athletic Humanoid Capabilities
| Spec | Unitree G1 (LATENT) | Unitree H1 | Figure 03 |
|---|---|---|---|
| Height | 1.32 m | 1.80 m | 1.70 m |
| Weight | 35 kg | 47 kg | 60 kg |
| DOF | 23-43 | 26 | 65+ |
| Battery Life | ~2 hrs | ~2 hrs | ~4 hrs (work) |
| Base Price | $13,500 | $90,000 | Not disclosed |
| Athletic Demo | Tennis (real-time) | Running / Parkour | Household tasks |
Who Built This: Galbot and Its Research Partners
Galbot Robotics (officially Beijing Galaxy General Robot Co., Ltd.) was founded in May 2023 by Wang He, a professor from Peking University. In under three years, the company has raised approximately $887 million across nine funding rounds. Its most recent round, a $362 million raise in early March 2026 led by China's National Integrated Circuit Industry Investment Fund (Big Fund Phase III), makes it the highest-valued unlisted humanoid robotics company in China. The company is reportedly eyeing a Hong Kong IPO.
Galbot's investor list reads like a who's who of Chinese industrial capital: CATL (the world's largest EV battery maker), SAIC Motor, CITIC Investment Holdings, China Development Bank, and Sinopec. The backing from CATL is notable because it signals battery-industry interest in humanoid robot power systems, not just autonomous vehicles.
• Galbot Robotics: Beijing-based, founded May 2023. ~$887M raised. Builds the Galbot G1 semi-humanoid and embodied AI platforms. Valued at $3B+.
• Tsinghua University: Co-developed the LATENT framework. One of China's top engineering and AI research institutions.
• Peking University: Founder Wang He's home institution. Contributed to the motion learning algorithms and sim-to-real pipeline.
• Unitree Robotics: Provided the G1 hardware platform ($13,500 base). Already used in dozens of research labs worldwide.
AI-generated image
Simulation environments running on GPU clusters train LATENT policies before deployment on physical robots.
What Athletic Robots Mean for the Industry
A tennis-playing robot might look like a party trick, but the underlying technology has serious industrial implications. The core problem LATENT solves, teaching robots complex physical behaviors from limited and imperfect data, is the same problem that factories, warehouses, and homes present. Real-world environments are messy. Workers don't move in textbook-perfect ways. Objects aren't always where they're supposed to be. If a robot can learn to play tennis from amateur footage shot in a cramped room, it can potentially learn to pack boxes, sort items, or assist with physical therapy from similarly imperfect demonstrations.
The choice of the Unitree G1 as the hardware platform matters, too. At $13,500 for the base model, the G1 is one of the most affordable humanoid robots available. Galbot's demonstration shows that research-grade athletic capability doesn't require a $100,000+ platform. That price point opens the door for university labs, smaller companies, and eventually consumer applications.
Broader Applications for the LATENT Framework
• Manufacturing: Learning assembly tasks from imperfect worker demonstrations rather than painstaking programming.
• Logistics: Warehouse robots that adapt to new packing patterns from a few hours of observation.
• Healthcare: Physical therapy assistants trained on therapist movement data.
• Sports and entertainment: Training partners, coaching tools, and athletic performance analysis.
China's robotics ecosystem is moving fast. In the past two weeks alone, Galbot closed its $362 million round, the Automation World 2026 expo in Seoul showcased Chinese humanoids alongside American and Korean competitors for the first time, and several Chinese manufacturers announced new production lines. The LATENT demo adds a research credibility layer to what has been primarily a commercial and manufacturing story.
Frequently Asked Questions
Can the robot actually beat a human at tennis?
Not yet. The Unitree G1 standing at 1.32 meters tall has significant physical limitations compared to a human player. It can sustain rallies and return shots at up to 15 m/s (about 34 mph), but competitive tennis serves regularly exceed 50 m/s (112 mph). The demo shows cooperative rallying, not competitive match play. Think of it as a capable training partner for casual hits, not a Wimbledon contender.
Why does this matter if it's not practical for factories?
The LATENT framework is designed to be generalizable. Tennis is the proof-of-concept because it's one of the hardest physical coordination tasks: fast reaction times, whole-body movement, and precise tool use (the racket). The researchers explicitly state the framework can extend to any domain where complete motion data is hard to get. Factory work, logistics, and household tasks all fit that description.
How much does the full system cost?
The Unitree G1 base model starts at $13,500. The EDU variant with dexterous hands and higher DOF ranges from $21,600 to $73,900 depending on configuration. Galbot has not disclosed pricing for the LATENT software separately. The motion capture data was collected with relatively standard equipment in a small space, keeping the research setup accessible compared to traditional sports robotics labs.
Has the research been peer-reviewed?
Not yet. The paper was posted to arXiv on March 16, 2026, as a preprint. It has not undergone formal peer review. The video demonstration provides supporting evidence, but independent validation and reproducibility by other labs will be important next steps.
What Comes After Tennis
Galbot's LATENT demo is a landmark for athletic humanoid robotics: the first time a bipedal robot has sustained real-time tennis rallies with a human using whole-body coordination learned from imperfect data. The 96% simulation success rate and the real-world video evidence suggest the framework works, though peer review and independent replication will tell the full story.
The bigger picture is about data efficiency. Most humanoid robot deployments today depend on either teleoperation (a human remote-controlling the robot to generate training data) or extensive simulation from scratch. LATENT shows a middle path: collect cheap, imperfect human motion data, extract primitive skills, and let domain randomization bridge the gap to reality. If that approach scales to other tasks, it could dramatically lower the barrier to deploying humanoid robots in unstructured environments.
The Bottom Line: A $13,500 robot just learned to play tennis from five hours of amateur footage. The tennis is impressive, but the real story is a training method that could make humanoid robots useful in places where perfect data doesn't exist.
Watch for Galbot's next move. With $887 million in the bank, a potential Hong Kong IPO on the horizon, and a research partnership with two of China's top universities, they have the resources to push LATENT beyond tennis courts. The question is whether other companies, including Figure AI, Tesla, and Unitree itself, will adopt similar imperfect-data approaches, or if they'll continue betting on teleoperation and massive simulation compute. Either way, the bar for what a humanoid robot can learn from minimal data just got raised.