Student Name: Karthik Shaji

Advisor: Dr. Yongxin Chen

Milestone: MS Thesis Proposal

Degree Program: Aerospace Engineering

Title: Towards Embedded Dynamic Semantic SLAM with 3D Gaussian Splatting

Abstract: Gaussian Splatting, an explicit method by which compact Gaussian primitives are used to represent scenes, has seen increasing interest lately over geometric methods and neural-implicit methods (i.e. NeRFs) of map representation due to its continuity, differentiability, unified geometry, and memory efficiency. Increasingly, they are seeing adoption for Simultaneous Localization and Mapping (SLAM) applications due to these inherent advantages. On the other hand, Semantic SLAM, where the objective is not only to perform map segmentation and factor graph localization (as in regular SLAM), but to also perform object pose and identity information, has become of increasing interest due to its ability to improve data association and loop closure. In this class of SLAM problems, there has been interest for the use of Gaussian Splatting due to its efficacy in regular SLAM problems. However, the current approaches have high memory requirements, and slow Frame Per Second (FPS) rates. This proposal aims to improve on the status quo for Semantic SLAM with Gaussian Splatting by optimizing it for embedded, space-borne applications. Notably, the standard NVIDIA AI chip used in research, the A100, has a runtime performance of 156 Tera Floating Point Operations Per Second (Flops). However, this chip has never been tested in a radiation environment. We aim to develop an algorithm capable of being run on NVIDIA Jetsons, which have a runtime performance of 475 GFlops, and which have been tested in Space, being viable for short-term space missions. To improve runtime, we propose leveraging Gaussian Process models, along with K-Nearest-Neighbors, to help with dynamic tracking and improve FPS performance by cutting memory overhead. The expected outcome is to make progress towards the development of embedded Semantic SLAM algorithms that can fly on intra-vehicular support missions or extra-terrestrial robotics applications.

Date and time: 2025-11-18, 1:00pm

Location: Online

Committee:
Dr. Yongxin Chen (advisor), School of Aerospace Engineering
Dr. Lu Gan, School of Aerospace Engineering
Dr. Yashwant Nakka, School of Aerospace Engineering
, 
, 
,