Katrina Ashton

I am a PhD student at the University of Pennsylvania, in the Department of Computer and Information Science, where I work on computer vision for robotics. I am co-advised by Prof. Kostas Daniilidis and Prof. Bernadette Bucher.

I have done interships with the Robotics and AI Institute and Seeing Machines. I received a Bachelor of Engineering (R&D, Hons.) in Mechatronics and a B.S. in Mathematical Modelling from the Australian National University, where I also worked as a research assistant with Prof. Jochen Trumpf and at the 3A Institute.

Email  /  Google Scholar

Research

I'm interested in computer vision, machine learning and robotics. In particular, I mainly work on scene representations for robotics, along with modular and interpretable policies for performing tasks using these scene representations.

HELIOS: Hierarchical Exploration for Language-grounded Interaction in Open Scenes

A zero-shot method for language-specified pick and place tasks by reasoning over a hierarchical scene representation with a global search policy to balance exploration and exploitation and an uncertainty-aware object score to allow us to only interact with objects after enough views have been captured to be confident of their class

Multimodal LLM Guided Exploration and Active Mapping using Fisher Information

An active Gaussian Splatting SLAM system with active exploration strategy which balances the accuracy of mapping and localization

Zero-shot Object-Centric Instruction Following: Integrating Foundation Models with Traditional Navigation

A zero-shot method to ground natural language instructions to a factor graph and a policy to use this for object-centric Vision-and-Language Navigation

Equivariant Filter for Feature-Based Homography Estimation for General Camera Motion

Homography estimation for arbitrary trajectories using the Equivariant Filter framework by exploiting the Lie group structure of SL(3)

Unordered Navigation to Multiple Semantic Targets in Novel Environments

Proposing an objective to extend an objectnav method to unordered multi-object navigation

An observer for infinite dimensional 3d surface reconstruction that converges in finite time

An observer for reconstructing the dense structure of scenes from visual or depth sensors that provably converges in finite time, demonstrated with real light-field camera data

Design and source code from Leonid Keselman's website