Physix FrontierPhysix Frontier · News Briefing Card (Leiphone · Oct 8, 2026)
Galbot and Tsinghua unveil LATENT for humanoid tennis rallying
KEY FACTS
- Teams from Galbot, Tsinghua University, Peking University and others released LATENT at IROS 2026, teaching a humanoid robot to learn tennis rallying.
- The study collected about 5 hours of motion data from 5 amateur players, covering only basic movements such as forehand, backhand and footwork.
- The method first trains a motion tracker compressed into a latent action space, then uses reinforcement learning to compose hitting strategies.
- Real-robot experiments used a Unitree G1, relying on more than 50 motion-capture cameras and a 19-meter by 15-meter motion-capture area.
- The simulated forehand success rate was 96.52%, while on the real robot the forehand reached 90.90% and the backhand 77.78%.
KEY DATA
96.52%Simulated forehand success rate
90.90%Real-robot forehand success rate
77.78%Real-robot backhand success rate
about 5 hoursMotion data duration
PHYSIX OBSERVATION
LATENT proves that imperfect human data can still give robots dynamic athletic ability, but the real robot still depends on an expensive motion-capture system. For the industry, this suggests sim-to-real need not pursue perfect simulation, but should actively inject uncertainty; for users, robot tennis remains at the laboratory stage and is still far from autonomous competition.
Source: Leiphone report