Organizers: Mustafa Mukadam, Arunkumar Byravan and Byron Boots
Website: https://sites.google.com/view/rss2018lair
Recent advances in machine learning techniques from the emergence of deep learning, and access to large amounts of data and powerful computing hardware have led to great strides in the state-of-the-art in robotics and artificial intelligence. In contrast to traditional approaches that are strongly model-based with priors and explicit structural constraints, these newer approaches tend to be data-driven and often do not explicitly model the underlying problem structure. As a consequence, while these approaches usually outperform their traditional counterparts on many robotics problems, achieving good generalization, task transfer and data-efficiency has been challenging. Combining the strengths of the two paradigms: (i) the flexibility of modern learning and inference techniques, and (ii) the domain knowledge and structural priors of traditional methods, should help bridge this gap.The goal of this workshop is to bring together researchers from robotics and machine learning to investigate, at the intersection of the two paradigms, techniques for structured learning and inference. Our notion of “structure” is very general. In the context of robot learning and inference this manifests in many ways: as a specific architecture of a (probabilistic) graphical model or (deep) network, an intermediate representation, a loss function, and so on. Some of the questions we hope to answer include: (i) How can we leverage structure to improve the state of the art in learning/inference models? (ii) What is the right mix between structure/priors/models and learning? and (iii) How can we establish benchmarks/baselines that show the effectiveness of using structure in learning/inference for robotics? A special emphasis will be on methods that tightly integrate structure with learning and are demonstrably applicable in the real-world, particularly on problems like autonomous navigation, manipulation, and field robotics.