Robust semantic representations for inferring human co-manipulation activities even with different demonstration styles.