Here we describe character animation production in the intelligent multimedia storytelling system, CONFUCIUS, which presents natural language stories as 3D animation, speech and non-speech audio. The modelling of animated characters in CONFUCIUS follows the H-Anim standard LOA1. It uses keyframing technology to perform the actions specified in story scripts, and has low-level autonomy such as path finding and collision reaction. Compared with other related virtual human modelling systems, CONFUCIUS’ character animation focuses on the language-to-humanoid animation process rather than considering human motion solely. We present the architecture of CONFUCIUS, and explain in detail design and implementation issues with illustratrative examples and the processes of how to get from semantics down to actual human motions.