Predicting humanoid motion has many interesting applications. This project was primarily motivated by the desire to create an artificial virtual game AI that can accurately target and evade a fully-articulated humanoid character. Due to the emergence of human motion interface devices, such as the Microsoft Kinect 3D camera system, players now have the ability to control their avatars through full body movement. To maintain a sense of immersion for the gamer, having artificial agents that can predict, learn, and adjust to the players motion input in a realistic manner is important. This project attempts to address this problem.
We build a supervised learner which can accurately predict the generalized mass motion of a humanoid. Specifically, we show that through use of support vector machines we can predict the movement of the center of mass (CM) for different rigid body chains of the character. Since we concern ourselves with large horizon windows, we do not attempt to predict the full-body dynamics of the character. Instead, we focus on the more general mass motion of the character that changes more slowly and predictably. For many applications, such as targeting the character with a projectile, this is sufficient.
To qualify our results we use a truncated Taylor series predictor. We demonstrate that not only is our SVM predictor more accurate than the Taylor series predictor, but also more scalable for large prediction horizons.
The blue character is the current character and the purple character is the future pose (0.5s ahead). The blue sphere is the future Center of Mass, the red sphere is our prediction, and the green sphere is the naïve Taylor series prediction.