Making the Coupled Gaussian Process Dynamical Model Modular and Scalable with Variational Approximations.

Area of Science:

• Robotics
• Machine Learning
• Biomechanics

Background:

• Human movement analysis often requires complex models.
• Reusable movement components (primitives) are essential for efficient human motion modeling.
• Existing models can be computationally intensive and lack modularity.

Purpose of the Study:

• To introduce a sparse, variational posterior approximation for the Coupled Gaussian Process Dynamical Model (CGPDM).
• To reduce CGPDM training time, enable modular reuse of learned dynamics, and facilitate compact storage of movement primitives (MPs).
• To apply this model to human movement primitive (MP) modeling for full-body movements.

Main Methods:

• Derivation of a sparse, variational posterior approximation for the CGPDM.
• Illustration of the approximation on synthetic (toy) data.
• Evaluation of model predictions against various MP models.
• Comparison of generated movements with human perceptual expectations.

Main Results:

• The variational CGPDM demonstrated superior performance in movement trajectory prediction compared to other MP models.
• Human observers found the generated movements nearly indistinguishable from natural movement recordings.
• The approximation achieved a very compact parameterization, crucial for large movement libraries.

Conclusions:

• The variational CGPDM offers an efficient and effective approach for modeling human movement primitives.
• The method successfully balances prediction accuracy, reusability, and storage compactness.
• This advancement has significant implications for robotics, animation, and understanding human motion.