Incremental learning of subtasks from unsegmented demonstration

Abstract

We propose to incrementally learn the segmentation of a demonstrated task into subtasks and the individual subtask policies themselves simultaneously. Previous robot learning from demonstration techniques have either learned the individual subtasks in isolation, combined known subtasks, or used knowledge of the overall task structure to perform segmentation. Our infinite mixture of experts approach instead automatically infers an appropriate partitioning (number of subtasks and assignment of data points to each one) directly from the data. We illustrate the applicability of our technique by learning a suitable set of subtasks from the demonstration of a finite-state machine robot soccer goal scorer.

Robot learning from demonstration by constructing skill trees

“…Finally, PDBV performs segmentation using Kinematic Centroid Segmentation, a heuristic specific to human-like kinematic motions in free-space, whereas CST uses a more generally applicable and statistically principled but potentially more expensive segmentation method. However, more recent work has used computationally expensive but principled statistical methods [Grollman andJenkins, 2010, Butterfield et al, 2010] to segment the data into multiple models as a way to avoid perceptual aliasing in the policy. Kulić et al [2009] use a principled, online and incremental method to perform segmentation, and use acquired motion primitives to build a hierarchy that can be used to improve later segmentation.…”

Learning grounded finite-state representations from unstructured demonstrations

“…The most closely related work to ours is that of Grollman et al [2010b], which addresses perceptual aliasing by using a nonparametric Bayesian model to infer a mixture of experts Figure 19: Segmentation of 8 demonstrations of the peg-in-hole task with repeated retries of insertions (top 2 rows) and 3 demonstrations of successful first-try insertions (bottom row). 36 from unsegmented demonstration data and then using multi-map regression to assign observed states to experts.…”

Learning to sequence movement primitives from demonstrations

“…In these cases also the switching behavior between movements is either deterministic or not learned at all [21]. The focus of our work instead is on incorporating several demonstrations with varying sequence orders into one graph model and learning the switching behavior between succeeding movements.…”