TaskNPoint: How to Teach Your Humanoid to Hit a Backhand in Minutes

A new training protocol called TaskNPoint can teach a Unitree G1 humanoid robot to hit a tennis backhand using short video demonstrations and less than an hour of computation on a single GPU, according to research published on arXiv [1]. The approach, detailed in a paper submitted June 24, 2026, rests on the observation that dynamic skills such as a tennis backhand are decided during a brief, critical portion of the motion — roughly 20cm of racket travel around ball contact [1]. Coordinating the entire body to get that narrow interaction window correct is what the system learns, rather than mimicking hours of broadcast footage [1]. TaskNPoint formalizes a coach-learner division of labor. A human coach supplies four inputs: a discrete set of skills, one demonstration per skill, identification of the interaction window, and the goal [1]. Training then proceeds inside a physically realistic simulation, which the authors say provides robustness to unmodeled events [1]. Randomized target sampling during training allows a single demonstration to generalize zero-shot to goal locations the robot has never seen [1]. The researchers tested the protocol on a Unitree G1 humanoid performing forehands and backhands against balls thrown by a person, kicking incoming soccer balls, and picking and placing boxes from novel locations [1]. In each case, learning succeeded from short human video demonstrations — a few seconds per skill — and required under an hour of training on one GPU, with no per-task reward tuning [3]. Prior work on teaching simulated characters tennis skills has relied on large-scale broadcast video collections and hierarchical policies that chain together extended rallies [6]. TaskNPoint departs from that data-hungry paradigm by focusing on the coach’s demonstration of a single skill and the critical interaction window, rather than harvesting thousands of hours of tournament footage [1][6]. The paper’s authors argue that the structural property of dynamic skills — that outcome hinges on a short, crucial trajectory segment — makes the coach-and-practice model both efficient and general [1]. The Unitree G1 hardware deployment confirmed that simulation-trained precision carried over to the physical robot without additional per-task adjustments [3].