Scaling Up Robot Reasoning Using Hierarchical Representations

Abstract: Hierarchical reasoning is an efficient symbolic reasoning approach that breaks complex problems into smaller, more manageable pieces, leading to faster and more optimal solutions. In this talk, I will demonstrate how introducing hierarchies enhance robot reasoning. For safety critical control under uncertainty, there is always a trade-off between safety (requiring faster response) and performance (requiring long horizon reasoning). We introduce a hierarchical long-short term safety framework that allows two safety control mechanisms to operate at different frequencies, thus advancing the safety-performance tradeoff. In data-driven prediction of other agents’ behaviors, incorporating hierarchy into neural network design improves generalizability and transferability across complex scenarios, such as vehicle trajectory prediction at intersections. For complex long-sequence planning, large linear temporal logic (LTL) specifications are challenging for users to specify and interpret, as well as for planners to reason over. We propose a new specification called hierarchical linear temporal logic (HLTL), which is not only more interpretable, but also easier to specify. Moreover, planning over HLTL is more efficient and scalable for multi-robot long-sequence planning. I will conclude the talk with future perspectives on leveraging hierarchy to enhance reasoning using large language models. 

Scaling Up Robot Reasoning Using Hierarchical Representations

Abstract: Hierarchical reasoning is an efficient symbolic reasoning approach that breaks complex problems into smaller, more manageable pieces, leading to faster and more optimal solutions. In this talk, I will demonstrate how introducing hierarchies enhance robot reasoning. For safety critical control under uncertainty, there is always a trade-off between safety (requiring faster response) and performance (requiring long horizon reasoning). We introduce a hierarchical long-short term safety framework that allows two safety control mechanisms to operate at different frequencies, thus advancing the safety-performance tradeoff. In data-driven prediction of other agents’ behaviors, incorporating hierarchy into neural network design improves generalizability and transferability across complex scenarios, such as vehicle trajectory prediction at intersections. For complex long-sequence planning, large linear temporal logic (LTL) specifications are challenging for users to specify and interpret, as well as for planners to reason over. We propose a new specification called hierarchical linear temporal logic (HLTL), which is not only more interpretable, but also easier to specify. Moreover, planning over HLTL is more efficient and scalable for multi-robot long-sequence planning. I will conclude the talk with future perspectives on leveraging hierarchy to enhance reasoning using large language models.