Inferring geometric constraints in human demonstrations