Roboticists have made significant strides in the last decade, enabling complex robot behaviors across diverse domains. Dominant approaches to robot intelligence typically fall into two categories. First, analytical methods offer reliability and robustness but require expertise and do not generalize beyond pre-specified scope. Second, learning-based approaches enable complex behavior without explicit design but demand significant resources and lack predictability. These categories are often viewed as competing alternatives due to their reciprocal advantages and drawbacks. However, they can be viewed as two ends of a broad spectrum.
This workshop seeks to explore a middle, structured approach that attempts to combine the benefits from either end of the spectrum. Such structured approaches can produce reliable and transparent behaviors with reduced compute and data requirements by incorporating structures from fields like dynamical systems, geometry, and physics. While promising, structured robot learning faces challenges in complex, open-world problems.
This workshop aims to bring together researchers working on structured learning topics. The goals are to identify advantages and limitations of structured learning, explore new directions for applying it to general robotic tasks, and foster collaborations. By bridging the gap between analytical and learning-based methods, structured learning offers a promising path forward in robotics research.
