SKID RAW: Skill Discovery From Raw Trajectories

Abstract

Integrating robots in complex everyday environments requires a multitude of problems to be solved. One crucial feature among those is to equip robots with a mechanism for teaching them a new task in an easy and natural way. When teaching tasks that involve sequences of different skills, with varying order and number of these skills, it is desirable to only demonstrate full task executions instead of all individual skills. For this purpose, we propose a novel approach that simultaneously learns to segment trajectories into reoccurring patterns and the skills to reconstruct these patterns from unlabelled demonstrations without further supervision. Moreover, the approach learns a skill conditioning that can be used to understand possible sequences of skills, a practical mechanism to be used in, for example, human-robot-interactions for a more intelligent and adaptive robot behaviour. The Bayesian and variational inference based approach is evaluated on synthetic and real human demonstrations with varying complexities and dimensionality, showing the successful learning of segmentations and skill libraries from unlabelled data.

MILD: Multimodal Interactive Latent Dynamics for Learning Human-Robot Interaction

“…III-B in [53]). The works closest to our approach [55], [68], [80] explore this key idea of learning how to segment demonstrations into underlying recurring segments or "skills" by decomposing a latent representation of the trajectories and subsequently learning the temporal relations between these skills to compose the observed trajectories. Similar approaches are further explored in literature using different skill representations such as the Options framework [75] or Autoregressive HMMs [44], [57].…”

Accelerating Robotic Reinforcement Learning via Parameterized Action Primitives

“…Many of these methods have not been extended beyond carefully engineered state based settings. More recently, research has focused on extracting useful skills from large offline datasets of interaction data ranging from unstructured interaction data [54], play [32,33] to demonstration data [2,39,43,44,48,50,58]. While these methods have been shown to be successful on certain tasks, the learned skills are only relevant for the environment they are trained on.…”

Learning Transferable Motor Skills with Hierarchical Latent Mixture Policies

“…A large body of work explores skills from the perspective of unsupervised segmentation of repeatable behaviours in temporal data (Niekum & Barto, 2011;Ranchod et al, 2015;Krüger et al, 2016;Lioutikov et al, 2017;Shiarlis et al, 2018;Kipf et al, 2019;Tanneberg et al, 2021). Other works investigate movement or motor primitives that can be selected or sequenced together to solve complex manipulation or locomotion tasks (Mülling et al, 2013;Rueckert et al, 2015;Lioutikov et al, 2015;Paraschos et al, 2018;Tosatto et al, 2021;Dalal et al, 2021).…”