KODA: Kinematics--Observation Data Alignment for Humanoid Learning from Human Egocentric Demonstration
KODA converts egocentric human demonstrations into whole-body vision-language-action (VLA) training data. It uses a reusable canonical trajectory to guide robot-view synthesis and generate controller-specific state-action supervision. This allows robots to learn from human videos more effectively, significantly improving success rates on real robot tasks when combined with existing robot data.