Standard motion tracking metrics miss the contact and stability problems that matter most to humans—HumanScore, trained on human preferences, better predicts what actually looks right in humanoid movement.
HumanTracker is a new benchmark for evaluating humanoid motion tracking that focuses on what humans actually perceive rather than just pose accuracy. It includes 153 hours of professional motion capture data and introduces HumanScore, a metric trained on human preferences that catches physical problems like foot skating and unstable balance that traditional metrics miss.