Variational Hierarchical Mixtures for Probabilistic Learning of Inverse Dynamics

Area of Science:

• Robotics
• Machine Learning
• Probabilistic Modeling

Background:

• Classical regression models in robotics face scalability or flexibility limitations.
• Need for computationally efficient and well-regularized probabilistic models in robotics is growing.

Purpose of the Study:

• To develop a novel probabilistic hierarchical modeling paradigm for robotics.
• To combine the flexibility of kernel machines with the scalability of automata.
• To introduce computationally efficient representations with inherent complexity regularization.

Main Methods:

• Probabilistic interpretations of local regression techniques approximating nonlinear functions.
• Bayesian nonparametrics to formulate flexible models with adaptive complexity.
• Two efficient variational inference techniques for learning hierarchical infinite local regression models.

Main Results:

• Demonstrated advantages in handling non-smooth functions and mitigating catastrophic forgetting.
• Enabled parameter sharing and facilitated fast predictions.
• Validated on large inverse dynamics datasets and real-world control scenarios.

Conclusions:

• Hierarchical infinite local regression models offer a powerful solution for complex robotics tasks.
• The proposed Bayesian nonparametric approach provides efficient, regularized, and adaptive probabilistic models.
• Successful application in real-world control scenarios highlights practical utility.