TaskNPoint trains humanoid robots to hit backhands in under an hour

A team of researchers led by Blake Werner at Caltech has developed TaskNPoint, a novel training protocol that dramatically accelerates how humanoid robots acquire dynamic skills. The key insight: unlike deep learning approaches that require thousands of demonstrations, TaskNPoint exploits a structural property of dynamic tasks—the outcome hinges on a short, critical portion of the trajectory. For a tennis backhand, this is the ~20cm racket travel around ball contact. The human coach provides just four inputs: a discrete set of skills, one demonstration per skill, identification of the interaction window, and the goal. Learning then happens in a physically realistic simulation environment, with randomized target sampling during training allowing a single demonstration to generalize zero-shot to unseen goal locations.

The approach was validated on a Unitree G1 humanoid capable of hitting forehands and backhands against balls thrown by a human, kicking incoming soccer balls, and picking/placing boxes from novel positions. Remarkably, training completes in under an hour on a single GPU, requiring no per-task reward tuning. The paper (arXiv:2606.26215) demonstrates that short human video demonstrations suffice, challenging the conventional wisdom that robots need massive datasets for complex motor skills. This coach-learner division of labor could fundamentally reshape how we teach robots practical physical tasks, making dynamic skill acquisition far more accessible.