Learning from demonstration using products of experts: Applications to manipulation and task prioritization

Probability distributions are key components of many learning from demonstration (LfD) approaches, with the spaces chosen to represent tasks playing a central role. Although robot configuration is defined by its joint angles, end-effector poses often best explained within several task spaces. In relevant learned independently and only combined at control level. This simplification implies problems that addressed in this work. We show fusion models different can be expressed as products experts (PoE), where probabilities multiplied renormalized so it becomes proper distribution angles. Multiple experiments presented jointly PoE framework significantly improves quality final model. The proposed approach particularly stands out when has learn hierarchical objectives arise requires prioritization sub-tasks (e.g. humanoid robot, keeping balance higher priority than reaching for an object). Since training model usually relies on contrastive divergence, which costly approximations affect performance, we propose alternative strategy using variational inference mixture approximations. particular, extended nullspace structure (PoENS), able recover secondary masked resolution higher-importance.