The work "Optimization Beyond the Convolution: Generalizing Spatial Relations with End-to-End Metric Learning", conducted in the lab of Prof. Wolfram Burgard, wins the Best Paper Award in Robot Vision at the International Conference on Robotics and Automation (ICRA 2018) in Brisbane, Australia. In the final round it leaves behind the two other nominated works presented by MIT and TU Munich. The jury finds that the paper tackles an important problem in robotic vision, namely modeling spatial relationships between objects and highlights its novelty as well as its solid validation.
The paper presents a deep neural network-based approach for learning representations that provide robots with the ability to understand arbitrary spatial relations between objects and to imitate spatial relations acquired from visual demonstrations with novel objects of various sizes and shapes. Imitating spatial relations is a desirable skill for future service robots that will perform manipulation tasks in human-centered environments, which are rich of arbitrary spatial relations and contain a large variety of different objects. For example, having learned how to place a book inside a drawer a robot should be able to generalize the same spatial relation to place a toy inside a basket. The presented work takes the next step towards better understanding of the perceived spatial structure of human environments and leverages the learned knowledge to imitate this structure in robot manipulation tasks. With this it significanly contributes to the BrainLinks-BrainTools project NeuroBots.
Paper: http://ais.informatik.uni-freiburg.de/publications/papers/jund18icra.pdf
Video: https://www.youtube.com/watch?v=XCSjZHi4cOw