A humanoid robot developed by Galbot Robotics is now playing tennis with humans in real time. Standing 4 feet tall, the robot reacts instantly without scripts or remote control, demonstrating full-body coordination and millisecond-level response times.
The system, called LATENT, runs on the Unitree G1 platform. It learns from fragmented human motion data-forehands, backhands, and side steps-collected from five players over five hours. Instead of relying on full-match recordings, the AI assembles these fragments into fluid sequences, enabling dynamic movement, shot placement, and recovery during live rallies.
Trained in simulation with variable physics, the robot adapts to real-world unpredictability. In tests, it achieved 96% success on forehand returns in simulation and sustains consistent rallies in live play. While not yet as fluid as human athletes, it shows early decision-making by placing shots strategically.
This advancement signals broader potential: the same AI framework could apply to sports like football or badminton, industrial automation, and search-and-rescue operations where complete motion data is unavailable. The technology brings us closer to robots training alongside or competing against professional athletes.