Chair: Maani Ghaffari

Abstract:
Robots require efficient and generalizable algorithms to obtain human-level autonomy in the future. Despite the rapid progress, modern robotic algorithms often require significant training data and do not generalize well to unseen scenarios. Symmetry, or “immunity to a possible change,” offers a promising pathway for designing efficient and robust algorithms suited to complex real-world environments.

In this dissertation, we explore incorporating symmetric structures to construct efficient and generalizable perception algorithms. We first propose a proprioceptive-invariant state estimation framework that operates in real time across various robotic platforms. Next, we present a learning-based proprioceptive contact estimator for legged robots, enabling state estimation on diverse terrains without physical contact sensors. We then develop a Lie algebraic neural network for equivariant representation learning on Lie algebras, which we apply to various tasks, including dynamics modeling, classification, and point cloud registration. Finally, we design convolutional neural networks equivariant to homography transformations, enabling perspective-equivariant representation learning. In summary, this dissertation highlights the advantages of using symmetry to address various robot perception challenges, resulting in efficient and generalizable perception algorithms.

Zoom link:
https://umich.zoom.us/j/99714074023?pwd=avVJXcBUfrC1NuUDwdoaKRkWl0w08F.1