Particle.news
Download on the App Store

Galbot Debuts Unitree Humanoid Rallying Tennis in Real Time With LATENT

The demo draws on a framework that composes brief, imperfect human motions into a decision-driven policy trained in randomized simulation.

Overview

  • The March 16 video shows a Unitree G1 returning shots against a human opponent using what appears to be a standard tennis racket.
  • In simulation, the LATENT system reached up to 96% success on forehands, and in the real-world test the robot sustained rallies and reacted to balls over 15 m/s.
  • Training relied on short motion fragments such as forehands, backhands, and side steps captured in a compact 3×5 meter space from five participants totaling about five hours.
  • To narrow the sim-to-real gap, the team randomized key physical parameters including mass, friction, and aerodynamics during simulation training.
  • Galbot worked with researchers from Tsinghua and Peking Universities, has released a not-yet–peer-reviewed paper, and says the approach could extend beyond tennis though validation remains limited to controlled demos.