Humanoid robot learns impressive tennis skills from imperfect human motion

Roboticists have struggled to get humanoid robots to effectively replicate athletic sports skills, such as those needed for tennis. These sports require highly dynamic motion, quick reactions, and high precision that robots are not usually equipped to handle. Past research attempted to use kinematic data and video-based extraction of human motion data, but these approaches were complex and often physically infeasible. Some robots have been trained to play sports like table tennis or football, but with limited agility and realism. Now, researchers in China have developed a system to teach humanoid robots how to play tennis far better than previous models. Their approach, described in a preprint paper on arXiv, uses imperfect, fragmentary human motion data that reduces the difficulty of the learning process.